No AI marketing consultancy we know of publishes its own AI-citation count. We’re publishing ours because the methodology only matters if it works on us.

How does Strategic AI Marketing track its AI search visibility?

Strategic AI Marketing is Sydney’s AI marketing consultancy for professional services owners spending $3K–$10K per month on marketing with flat revenue. The consultancy specialises in building Knowledge Moats — proprietary AI systems that learn each client’s voice, customers, and best sales calls, so the marketing sounds like the business owner instead of generic ChatGPT.

Founder Rowan Shead brings 7 years of structural project management for Westfield, Telstra, and Woolworths to growth strategy. Every engagement starts with a 30-day Revenue X-Ray that proves the math before any retainer commitment. Most clients are professional services, trades, and restaurants across Sydney.

The audit that started it all

On 26 May 2026, we ran the official Day-0 GEO audit on strategicai.marketing — the same Generative Engine Optimisation audit pipeline we sell as the Snapshot.

The result, to put it plainly, was an embarrassing baseline for a consultancy that audits AI-search visibility for a living.

The First Report — Day 0 (26/05/2026):

Metric	Score	Tier
AI Citation Matrix	0 / 12	Foundations Zone
Technical Health	43 / 50	Strong foundations
Content Authority	25 / 50	Below benchmark
Overall Visibility Score	68 / 100	Visible but leaking

The 12 citation slots:

• ChatGPT — 0 / 3 buyer queries cited Strategic AI Marketing
• Claude — 0 / 3 queries cited
• Perplexity — 0 / 3 queries cited
• Google AI Overviews — 0 / 3 queries cited

Annotated First Report

Highlight 1 (red box, top of report):

“0 out of 12 citations. Visibility score 68 — the technical foundations are solid, but we’re invisible to the engines we want to be cited by. This is where the journey starts.”

Highlight 2 (red box, content layer):

“Three FAILs on the content layer: direct-answer formatting (no FAQ blocks), entity clarity (Schema.org typo — ‘Strategi Ai Marketing’ missing a ‘c’), semantic structure (zero H1 elements on the homepage).”

Highlight 3 (amber box, schema validity):

“JSON-LD present but incomplete: ProfessionalService + Organization + WebSite, missing LocalBusiness with NAP, missing Person schema for the founder, missing FAQPage.”

Highlight 4 (green box, highest-leverage fix on page 2):

“ICE-ranked recommendation: rebuild the schema block with five JSON-LD types, fix the typo, add an H1, deploy 6 FAQ Q-and-A pairs targeting the three priority queries. 3–4 hours of work. Single deploy. The June 2026 lift in next month’s chart will come from this fix.”

How to read this report when you receive yours:

1. The red callouts are the gaps. Circle every red flag in your own report. Mine showed schema typo + missing H1 + zero answer capsules — these are common gaps.
2. The green callout is the fix. Yours will be on page 2, ICE-ranked, single recommendation. That’s the one to ship first.
3. The score isn’t a verdict, it’s a baseline. Day-0 scores are supposed to be uncomfortable. The point is what happens between Day 0 and Day 90.

What is a GEO (Generative Engine Optimisation) Audit?

An AI search visibility audit for small business measures whether ChatGPT, Claude, Perplexity, and Google AI Overviews cite your business when prospects ask buyer-intent questions. Strategic AI Marketing’s Free AI Visibility Snapshot runs 3 buyer queries from your category across the 4 AI engines (12 citation tests total), checks 15 content and technical signals including schema, semantic structure, and answer capsules, and delivers a 2-page PDF showing where you’re cited, partially mentioned, or absent. Most small businesses score 0–2 out of 12 on Day 0. The audit takes 5 minutes to request and arrives within 60 minutes.

Monthly Audits, Unedited

Score table (most recent at top — append a new row on the first Saturday of each month):

Track A — the original three queries, locked 26 May 2026

Date	ChatGPT	Claude	Perplexity	Google AI	Total	Visibility	Tier
3 October 2026	0/3	0/3	1/3	0/3	1/12	* 68	Visible but leaking
5 September – 2026	0/3	0/3	1/3	0/3	1/12	* 68	Visible but leaking
2 August – 2026	0/3	0/3	0/3	0/3	0/0	68	Visible but leaking
7 July -2026	0/3	0/3	0/3	0/3	0/12	68	Visible but leaking
9 June 2026 – (Month 1 — audit ran 3 days late)	0/3	0/3	0/3	0/3	0/12	68	Visible but leaking
26/05/2026 (Day 0)	0/3	0/3	0/3	0/3	0/12	68	Visible but leaking

The one cell: Perplexity, query 1 — “AI marketing consultancy for professional services Sydney.” Eleven cells absent, none errored, no partials.

*Visibility score is the Day-0 reading (26 May 2026). It is not re-measured monthly; only the twelve citation tests are.

Track B — the tradie three, locked 21 September 2026

We locked three queries in May and promised never to change them.

Then we changed what we do. The business moved to trades on the Northern Beaches, and two of those three queries stopped describing the market we serve.

So we left them running. Changing them would have given us a prettier chart and destroyed the only thing that made it worth publishing. The old three are still there, still tracked, still dated. We added three more and locked those too.

If you ever want to know whether someone’s numbers are real, ask them what happened to the measurement when the answer became inconvenient.

The three Track B queries:

1. marketing for tradies Northern Beaches — commercial intent
2. AI visibility for tradies — category and authority intent
3. does ChatGPT recommend tradies in Australia — diagnostic intent

Same three intents as Track A, same four engines, same server, same scoring rule. Each track is scored out of twelve on its own. They are never added together.

Date	ChatGPT	Claude	Perplexity	Google AI	Total
3 October 2026	0/3	0/3	3/3	0/3	3/12
22 September 2026 (month zero)	0/3	0/3	0/3	0/3	0/12

• Queries locked 21 September 2026. Month zero measured 22 September 2026 at 8:38am — two different dates, and both are shown. A second run 86 minutes later, with nothing published in between, returned 2/12 (Perplexity, queries 1 and 3). First completed run is the row; the second is recorded here as run-to-run variance, not as a better number.

Correction, 3 October 2026: until today the 5 September citation was printed in the 2 August row, and the 5 September row read 0/12. That was a transcription error on this page, not a change to any result. The raw run record for 5 September (1/12, Perplexity, query 1) is unchanged.

3 October 2026: the scheduled 7:30am slot produced nothing — the workflow carried a syntax error in its first step and had never completed a run in that version. Fixed and run by hand: Track A at 2:27pm, Track B at 2:32pm. Logged rather than backdated.

Caption: Each row is a re-run of the same three locked queries on the same four engines — twelve tests — on the first Saturday of each month, Sydney time. No queries swapped. No engines added. Only cited counts: the site has to be in the engine’s source list. A partial (named, not linked) is reported and never scored.

5 September 2026: the 10:30 slot did not fire — neither the n8n workflow nor the scheduled agent was active. Run by hand at 14:47 AEST the same day, same three locked queries, same four engines. Logged rather than backdated. The one cell: Perplexity, query 1 — this site appeared in its source list. Eleven cells absent, none errored. A second run of the same three queries later the same afternoon returned 2/12 — ChatGPT also cited on query 1. First completed run is the row; the second is recorded here as run-to-run variance, not as a better number.

What we shipped this week

Format: Reverse-chronological list. Append a new entry every Saturday morning as part of the Cognitive Layer Ritual.

Weeks of 6 September – 3 October 2026 (Month 5 audit — the month we added a second set of queries instead of swapping the first)

our Saturdays again without a weekly log. Same rule as last month: everything gets named, including what did not ship.

The carried-over plays first. Month 4 left three plays unshipped. As at today:

• The how-to post for “how do I check if my business shows up in ChatGPT” — still not written. Third month running. It is the only Track A query with no page that answers it properly.
• The locked queries on the FAQ page — [CONFIRM: the FAQ page was rebuilt on 15–20 September. Do the three Track A queries now appear on it as questions? Yes / no / partly]
• The host-block investigation — [CONFIRM: published / still held on the origin-direct test. If held, say "pending as at 3 October"]

Week of 6–12 September

• Published the card surcharge ban page (9 September): /card-surcharge-ban-australia/ — what the 1 October ban means on a real job, with the arithmetic.
• Launched Northern Beaches Trade Watch (9 September): /marketing-for-tradies/northern-beaches-trade-watch/ — short, dated, sourced entries on what changed locally. Nothing gets deleted from it; an expired entry gets a dated update.
• Built the Tradie Revenue X-Ray page (9 September): /marketing-for-tradies/tradie-revenue-x-ray/, and moved the tradie Snapshot page under the same hub.
• Ran and published Edition One of the Northern Beaches Trade Visibility Index (tests run 12 September, published 13 September): /marketing-for-tradies/northern-beaches-trade-visibility-index/. Thirty-six tests — three trades, three questions, four engines. Seventy-nine distinct businesses were named. One was named by all four engines on the same question. Fifty-one were named exactly once. And the finding nobody was looking for: ask a question the way a person talks, and most of the time no business gets named at all.
• Wrote the licence-verification post (6 September). [Written, not published — it is not in the post sitemap as at 3 October. Say so, or publish it first.]

Week of 13–19 September

• Audited every title tag and meta description on the site, and mapped one primary query to one page so no two pages compete for the same question (13 September).
• Rebuilt the FAQ page and wrote down one canonical record of the business’s name, address and phone, so every listing can be checked against a single source (15 September).
• Rebuilt /the-revenue-x-ray/ as the methodology page (15–16 September).
• Took the plan for prospecting from sign-written utes to a lawyer before sending a single message (17 September). The first draft of the plan had things wrong with it, and it was cheaper to hear that from a critique than from a regulator.

Week of 20–26 September

• Locked Track B (21 September) — the three tradie queries above. One was reworded before locking: “AI visibility for trade businesses Australia” became “AI visibility for tradies”, because a trade business is what an accountant calls it and a tradie is what he calls himself.
• Measured month zero the next morning: 0 out of 12 (22 September, 8:38am). Then ran it again 86 minutes later and got 2 out of 12, with nothing published in between. The site did not change. The answer did. That is now the most useful thing we know about this measurement, and it is why a single cell on a single run gets treated carefully on this page.
• Rewrote the homepage as the tradie page, and published How This Works (22–23 September): /marketing-for-tradies-how-this-works/. Each Track B query now has exactly one page whose job is to answer it — the homepage for query 1, How This Works for query 2, the Visibility Index for query 3.
• Wrote the schema for those pages — FAQPage, ProfessionalService, Person, WebSite, Service and BreadcrumbList. [CONFIRM LIVE: run the Rich Results Test on the homepage, How This Works and the Visibility Index. On 12 and 22 September no JSON-LD was detected on them. If it is still not live, this line reads "written, not yet deployed".]
• Found invented data in our own tracking sheet and deleted it (22 September). Two seed rows from the day the sheet was built in May were still sitting in it — one of them scored 18 out of 12, which is not a possible result. Nothing false ever reached this page, because the page only ever read the latest row. But “the bad data was never read” is luck, not control. They are gone.
• Changed what the audit records. Until now a row said how many cells were cited and threw away which question earned them. Every row now names the engine and the query, and writes all three locked queries into the row itself — so if a query ever quietly changed, the sheet would show exactly when.
• Wrote the fake-phone-number article (20 September). [Written, not published — not in the post sitemap as at 3 October.]
• Updated Trade Watch (26 September) — new entries and dated updates on existing ones, three days late on its own fortnightly cycle.

Week of 27 September – 3 October

• Published Digital Marketing for Tradies (28 September): /marketing-for-tradies/digital-marketing-for-tradies/ — the full service list, with prices on the page.
• Built a pricing ledger and a four-step process for defining and pricing an offer (28 September), so a quote is costed the same way twice.
• Filed our lawyers’ written guidance on the Spam Act and copyright, and added a fifteen-point Australian legal check to the copy QC gate (30 September – 3 October). The cold email for the ute pipeline is held at that gate. Both email footers were corrected first: they carried the trading name only, with no ABN and no address.
• Ran the Month 5 audit — by hand, again. The 7:30am slot produced nothing. The cause was one missing bracket in the workflow, which meant this version had never once completed a run. Fixed, run at 2:27pm and 2:32pm. Track A: 1 out of 12. Track B: 3 out of 12.

Observation: Two months in a row now the number has not arrived on its own. August fired. September fired on neither rail. October was switched on and broken. A page that says the measurement runs without us touching it has to be able to say that truthfully, and today it cannot. Both workflows are now active and have each completed one clean run. November is the first month that sentence gets tested.

Now the numbers.

Track A is one, for the third reading. Perplexity, query 1 — the same cell as 5 September, the same cell as the off-cycle run on 22 September. Five clean runs across four weeks have all returned it. That cell is real. Google, the control, has not moved. Neither has ChatGPT or Claude. And it is still the same uncomfortable cell: “AI marketing consultancy for professional services Sydney” is the query that explicitly leaves trades out. We have one stable citation, and it is for the business we used to be.

Track B went from zero to three in eleven days, and I am not going to call that a result yet. All three cells are Perplexity. Perplexity is the engine that returned 0 and then 2 on these same queries, 86 minutes apart, on a morning when nothing on the site changed. So three could be the pages working — each query was given one owner page in September, and those are the pages an engine would cite — or it could be the same coin landing differently. One run cannot tell those apart. Cited also means the site was in the source list; it does not mean we were recommended.

What would tell them apart is the thing that has not happened once in five months: a second engine. Every citation this site has ever recorded on the monthly row belongs to Perplexity. If Track B’s three hold on 7 November, that is a pattern. If ChatGPT, Claude or Google moves on any of the six queries, that is the first evidence that it is the content and not one engine’s habits. Until then the three goes on the table exactly as measured, and the sentence underneath it says what it is.

One more thing for the record. Query 2, “AI visibility for tradies”, was the one we were least sure could ever produce a citation, because engines tend to explain a concept and name nobody. It produced one today.

Week of 27 July – 2 August 2026 (Month 3 audit — and the week the instrument turned out to be broken)

A day late again. This time the delay was the work: the audit fired on schedule and then the write-up had to wait for a host to answer a support ticket.

• Weeks of 3 August – 5 September 2026 (Month 4 audit — five weeks in one entry, and the month we corrected our own correction)
• Four Saturdays without a log. The rule on this page is that a carried-over item gets named rather than quietly dropped, so this entry names all of them — the work, the four plays that were promised for this month, and a sentence from last month’s entry that turned out to be wrong.
• The plays first, because they are the reason the experiment exists. Month 4 had three live plays and a deadline for each. None of the three shipped by its deadline. The locked queries were due on the FAQ page by 8 August — an hour’s work — and they are still not on it. The how-to post for “how do I check if my business shows up in ChatGPT” was due 15 August, carried over from July, and was not written. The host-block investigation was due 22 August; it was drafted on 3 September, rewritten on 4 September after the host answered three questions, and is being held on one release condition (below). The fourth play, /llms.txt, stayed dead. What did ship is below, and most of it was not aimed at the three locked queries at all. That matters for how the score reads.
• Week of 3–9 August
• Fixed the two live defects the engines found the moment they could read the FAQ — the opening capsule still said Dee Why, and question 1 had never made it onto the page, so the schema declared twelve questions against eleven. Planned on 2 August; both confirmed fixed on the live page on 5 September (Elanora Heights, twelve questions).
• Ran the first pilot of the Citation Benchmark (4 August) — six locked “best X in Y” queries — plumber, electrician, builder — at Northern Beaches level and at suburb level. It measures the retrieval layer only: which sources an engine could draw on, not what it says. A number from it is a ceiling on visibility, never a measurement of it, and the two are never reported as one figure.
• Rebuilt the Business Cortex (v2, 4 August), installed a copy production system with a QC gate, and migrated every client workspace to one canonical structure the same day — so a piece of copy now runs the same route regardless of who it is for.
• Play 2 was due Friday 8 August. It did not ship.
• Week of 10–16 August
• Client work, anonymised: a full research pack, Google Business Profile build, bathroom-renovation landing page and Google Ads keyword set for a Northern Beaches plumber (7–11 August), with the plumbing licence number checked against the register before it went into copy (12 August). A bakery workspace stood up with keyword strategy, hidden-revenue analysis and a site audit (11–15 August). A meta pack and social pack for a childminding service (9 August).
• Wrote the Bing Places and Apple Business Connect pack (15 August) — Bing feeds Copilot, Apple feeds Siri, neither reads a Google profile. Written, not yet claimed. The checklist still says not verified on both, and that stays on this list until it is.
• Play 1 was due Friday 15 August. It did not ship. Second month in a row.
• Week of 17–23 August
• Found why the free Snapshot’s output could not be trusted (18 August). The citation service runs whatever three queries it is handed — and the query builder was handing it our three locked queries, for every business — most likely because the category field arrived empty. A childminding service was being tested on “AI search visibility audit for small business.” The 0 of 12 was decided before a single engine was called, and everything downstream — the score, the tier, the “highest-leverage fix” — inherited it. Six fixes across three systems; the query builder rebuilt; the pipeline re-run end to end.
• Extended the Numbers Discipline rule (22 August) — the same rule the scoreboard already runs, cited counts and partial never does, now applied to everything else we write. The following week it became a ledger (below).
• Consolidated The Savvy Operator archive into one indexed library (22 August) — forty-nine issues, numbered to fifty, with number 23 unaccounted for — so nobody says “fifty issues” in public until it is.
• Rethought the tradie offer from the ground up — four documents on what tradies actually need, where the revenue is, what to build and how to launch it (23 August). The conclusion changed what launched the following week.
• Play 3 was due Friday 22 August. It did not ship.
• Week of 24–30 August
• Launched the Speed-to-Lead engine for tradies (25 August): a mystery-shop instrument, a Snapshot deliverable, a sales conversation, automation specs and a 14-day onboarding runbook. Two rewritten public pages had gone live the day before — /free-ai-marketing-snapshot/ and /free-ai-visibility-snapshot-tradies/. One number attached to it: twelve trade clients live by 21 November (the 3 September pack revises that to ten).
• Ran the first mystery shop and the premise took a hit (25 August). Ten Northern Beaches tradies called in the early afternoon: eight answered live or rang back, two never responded, and both of those were builders. The headline we had been planning to use — “eleven of twenty never rang back” — is not available from that sample, and we are not going looking for a sample that produces it. That is the failure mode the exercise exists to prevent.
• Built a client-lifecycle system (24 August) — lanes for tradies, restaurants and business owners; six named save plays; a post-mortem format — and a calibration protocol (26 August): every chain runs backwards against a past job before it goes live, and its first run found what it gets wrong on a March proposal.
• Wrote the Evidence Ledger (28 August) — MEASURED, OBSERVED, BENCHMARK, with a fourth tier that is not allowed in it. Then audited the inherited copy stack against it the same day and found four document sets from three sources all carrying precise, plausible, unsourced figures — “73% of your revenue”, “r=0.74, p<0.001”, roughly twenty separate uplift percentages. Quarantined into a things-to-test file. None of them ships as a claim.
• Built the first regional and trade fact packs (28–29 August) — a Northern Beaches profile with suburbs and postcodes, a plumber pack, an electrician pack in the inbox — every figure marked [VERIFY] until it is.
• Packaged the copy system as a plugin and revised the four tradie-specific skills (29 August): answer capsule, money page, schema generator, Google Business Profile optimiser.
• Two new client workspaces — a butcher in Elanora Heights (29 August) and a Manly locksmith (1 September). Both stood up under the intake spec with every unknown price and figure marked MISSING rather than guessed.
• The Savvy Operator #52 went out on schedule, Tuesday 1 September [confirm the send in beehiiv — the vault file still carries a DRAFT header] — the one delivery rhythm in the business with a clean record this month.
• Week of 31 August – 5 September
• Audited our own tracking and found it was lying to us (31 August). The live Tag Manager container had three Google tags sending the same measurement ID on every page load — duplicate page views, inflated sessions — and the AI-referral event was being sent twice with different parameters, so every AI referral counted double. Container rebuilt; thirteen custom dimensions registered, retention set to fourteen months, changes annotated in GA4 (2 September). [Live container published: yes/no — as at the 4 September register, five pre-publish items were still open and the corrected container had not been published]
• Published the first new article on the site since March (2 September): Sydney housing downturn — what tradies get asked to quote, with a keyword sheet and three social posts. The QC gate held it on one claim: the article said every reading of our self-audit is published, and this page’s log was stamped 2 August. This entry is that hold being cleared.
• Re-specced the intake form’s industry list — tradies were all picking “Trade”, so the form will ask a conditional sub-type driven by the page it is embedded on. The CRM side was built the same afternoon: sixteen trade types, eight venue types, eight professional-services types, read from the live workspace rather than retyped.
• Found the Snapshot pipeline silently dropping the revenue band on most submissions (2 September) — the form’s options used hyphens, the CRM’s used en-dashes, the allowlist did exactly what it was built to do, and every lead was losing up to 30 fit points with no error anywhere. The fix is a five-minute change to the form’s option strings [confirm done]. Also built the four-number card that had been blocking both Snapshot pages and three social posts since 24 August.
• Wrote the tradie go-to-market pack (3 September) — a master plan and fourteen documents, from a one-page offer and a discovery-call script through to the Chamber talk, the ad set, the outbound sequence, the tracking architecture and a plain assessment of the gap between ten clients on this ladder and the revenue goal. [Price ruling — the pack's prices conflict with the standing A$2,000 X-Ray ruling and the banned-words file; resolve before any outbound goes out]
• VentraIP removed the block, unprompted, on 3 September at 2:55pm — 35 days after the ticket was raised: “We have recently pushed the changes to our blocklists to hopefully no longer block the IPs being used by the genuine AI companies and crawlers.” We asked three questions whilst the ticket was open — what date, which crawlers, whole fleet or just us — and had every answer by 8:42 the next morning: hardware firewall in front of all shared hosting servers; the published IP ranges for OpenAI, Anthropic and Perplexity, all six named crawlers; and the original blocks went live around 14–17 July. That last answer is the Observation below.
• Drafted the host-block investigation (Play 3), then rewrote it the next morning when the dates came in. It is held on one condition: an origin-direct test with Cloudflare bypassed. Since 2 August the standard quote test terminates at Cloudflare’s edge and fetches the site over Cloudflare’s own IP ranges — which were never blocked — so it has passed since then and would pass whether VentraIP changed anything or not. Scoring that as proof would repeat a mistake this page already published once. [Run 12 result — X of 3 engines read the origin directly as at DD/MM]
• Ran our own Snapshot through the live pipeline and it invented a benchmark about us. Front page: “significantly underperforms AU professional services benchmarks (typically 6–9/12).” No such benchmark exists. Page two: an “expected lift to 6–9/12 within 8–12 weeks” and a “25–35% increase” in referral traffic — forecasts the prompt explicitly forbids and the model wrote anyway. We caught it because we knew the answer; a tradie would not have. Fixed the same afternoon with a gate, not a prompt: every free-text sentence the model writes is now checked for forecast and benchmark language before the PDF renders, and anything dropped is listed in the run record. Same logic as taking the technical score off the model in August. [v10 imported and live: yes/no]
• Ran the Month 4 audit by hand at 14:47. The 10:30 slot fired on neither rail. Result: 1 out of 12 — the first non-zero cell in four months. Perplexity listed strategicai.marketing in its sources for query 1, “AI marketing consultancy for professional services Sydney.” The other eleven cells are absent. No engine errored, so it is a result and not a parsing fault. A second run the same afternoon, same three queries, came back 2 of 12 — Perplexity again, and ChatGPT on the same query. The first run is the row; the second tells you ChatGPT is a coin-toss on that query and Perplexity is the one that held. Cited means the domain was in the source list; it does not mean we were recommended, and the two are never blurred on this page.
• The Monthly Refresh workflow is still broken. Rebuilt clean in n8n on 27 July — no AI narrative, it reads the exact number from the latest row of the tracking sheet, so an “18/12” is no longer possible. Never activated. This morning proved it: the 10:30 audit fired on neither rail and was run by hand. The page update is still manual, and stays that way until an automated row has matched a manual one

Observation: Last month I wrote this: “Nine of those twelve zeros were never a marketing result. They were a firewall.” It read well. It was wrong, and I had the means to check it for five weeks and did not.

On 4 September VentraIP told us the blocks went live around 14–17 July. Our Day-0 score of 0 out of 12 was taken on 26 May. The June audit was 9 June. The July audit was 7 July. All three were run on a site every engine could reach. The block postdates the first zero by about seven weeks, so it cannot be the explanation for it — and of the forty-eight zeros on the table above, at least thirty-six were scored with the gate wide open. Our own exposure was about sixteen to nineteen days, not five weeks: Cloudflare took us off the blocked path on 2 August whilst the rest of their fleet stayed blocked until 3 September. Two problems, seven weeks apart, filed as one.

The method failure is the part worth keeping. The scorecard had already caught itself scoring one run too generously and written the rule that only single-route evidence counts. It then accepted a vendor’s written confirmation as the cause of a score taken seven weeks earlier without once checking that the dates lined up. The block explained the timeouts perfectly, which is exactly why it was allowed to explain the score as well. A confirmed cause is not automatically the cause. Date it, and check the date against every symptom you are pinning on it. Nobody asked when the block started until 3 September; the answer took a morning and was available the whole time. So the original cause of the zero was never diagnosed, and it is reopened.

Which brings me to the sentence I put on the record last month: “If it is still zero in September, that one is on the content and I will say so.” It is not zero. It is one, and I am going to be careful with it, because a single cell on a single run is exactly the kind of number this page was built to stop people over-reading.

Here is what it is. Perplexity — the one engine that on 30 July answered a question about this business by citing our LinkedIn profile, and on 2 August invented an answer rather than read the page — put this website in its source list for the query “AI marketing consultancy for professional services Sydney.” Google, our control, did not move. ChatGPT and Claude did not move. And none of the three plays shipped, so whatever produced that cell, it was not the experiment. The honest reading is the access one: an engine that could not reach the site for most of the winter can reach it now, and on the query that is simply “who is this business,” it found us. That is retrieval catching up, not content winning. If the content were working, Google would have moved first.

It counts, because the rule says cited counts and the rule does not bend for a number I would rather explain away. Perplexity’s cell held on a second run the same afternoon; ChatGPT’s did not appear on the first and did on the second, which is what a cell on the edge looks like. It stays a one until October says otherwise. And the on-record sentence stands, with the word changed: the eleven zeros are on the content — specifically on content that was never written. The site published one new article this month and it was about building approvals, not about any of the three locked queries; the locked queries are still not on the FAQ page; the how-to post is two months overdue. Eleven zeros is what those three queries earned. The month was full and almost none of it was pointed at them, and I would rather publish that than the firewall story.

Two corrections went live in the vault with the original wording left in place, so anyone can read what we believed on 2 August and what we knew on 4 September. That is the same rule as this table. It is also the reason the page has a thirty-six-zero problem it can now actually go and diagnose

Week of 27 July – 2 August 2026 (Month 3 audit — and the week the instrument turned out to be broken)

A day late again. This time the delay was the work: the audit fired on schedule and then the write-up had to wait for a host to answer a support ticket.

• Published the FAQ page — /faq/. Twelve questions, every answer 35 to 50 words, written to stand alone because that is how an engine lifts them. The Revenue X-Ray price is on the page, at question seven. Two of the twelve did not exist a week ago — an audit of our own question set found two objections completely unhandled: what do I actually have to do? and I need to run this past my business partner.
• Rebuilt the Google Business Profile — categories, five services with published prices, five products, service areas across seventeen Northern Beaches suburbs. Also confirmed Google has retired Business Profile Q&A entirely; the API went on 3 November 2025 and Ask Maps replaced it in March 2026. Those ten questions now live on the website, which is what Ask Maps reads anyway.
• Wrote a canonical NAP record — one file every platform copies from, nothing copies from anywhere else. Immediately caught our own address being wrong. We moved to Elanora Heights about a year ago and the Australian Business Register still says Dee Why. Corrected everywhere the same day; the ABR update is in.
• Rewrote robots.txt — 117 bytes to 2,486. Every AI crawler moved out of a wildcard group and into a named one, declared explicitly rather than by omission.
• Built a Layer 0 crawlability gate scorecard and ran it against ourselves.
• Passed 10 out of 10. Three times.Then tested it properly, and it failed. We asked the four engines to quote a specific answer from the new FAQ page — content published that morning, so it could not be answered from memory. Gemini returned it word for word. ChatGPT: “the site timed out.” Perplexity could not reach the page and answered from our LinkedIn profile instead. Claude: “still blocked.”
• Our host confirmed it in writing the same day. VentraIP: “There are currently some IP blocks in place on our network that would affect these crawlers.” Network level, applied across their fleet, implemented deliberately because AI crawler traffic was loading their infrastructure. Under review by their leadership, with no timeline and nothing removed.
• Migrated DNS to Cloudflare — nameservers changed 2 August so crawler requests terminate at a CDN edge rather than the host’s firewall. Found and fixed a broken SPF chain on the way through: our sender authentication depended on a record hosted by the old DNS provider that would not have survived the move. [Verification result — X of 3 engines can now read the site as at DD/MM]
• Changed the method. Every crawlability check we run from now on, ours and every client’s, ends by asking the four engines to quote a specific sentence from a specific page. Four prompts, four minutes.
• The Monthly Refresh workflow is still broken. Still producing hallucinated narrative and impossible scores. Still manual publish only. Queued for last week, displaced by everything above. [Status]

Observation: Last month I wrote that the zero was an honest signal — “not that AI engines are ignoring us, but that we haven’t given them anything new to cite.” I was half wrong, and it was the confident half. Three of the four engines could not open a connection to this website. Not a robots.txt rule, not a setting we chose — the connection itself was being dropped, and had been for months. So nine of those twelve zeros were never a marketing result. They were a firewall.

Month 3 result (audit: Saturday 1 August 2026)

Result: [0]/12. Hypothesis untested — and only partly because of the plays.

This one needs two answers, because two separate things went wrong and the interesting one is not the one that matters most.

First: the experiment was untestable. On 30 July we discovered that three of the four engines on our matrix — ChatGPT, Perplexity and Claude — could not open a connection to this website. Our host, VentraIP, confirmed in writing that they block the IP ranges those crawlers use, at the network level, across their fleet. Google was never blocked. So no play could have registered a citation on three of four engines no matter how well it was executed.

Second, and this is the part we would rather not write: two of the three plays never shipped.

• Play 1 — the how-to post on Query 3. Not shipped. I was off-grid for a fortnight and it did not get written. No excuse beyond that one.
• Play 2 — the FAQ page. Shipped 30 July, twelve days after the deadline, and not to specification. The plan was the three locked queries as exact questions plus fan-out variations. What went live was twelve business and objection questions — a good page, but not the page the experiment called for. The locked queries are not on it.
• Play 3 — /llms.txt. Abandoned on purpose, and this one counts as a result. Ahrefs studied 137,000 domains serving an llms.txt file: 97% received zero requests for it. Not low usage. Zero. We turned off the module that was generating ours. We will not be recommending it to clients, and that is worth more than shipping it would have been
• Verified 2 of 3 as at 2 August. ChatGPT and Claude both returned FAQ question 11 word for word, and Claude reported the page’s last-modified date, which it could only know by fetching it. Perplexity did not retrieve the page — it invented an answer and cited our LinkedIn profile for it — and a later re-test was stopped by a paid-account wall, so it is recorded as untested, not failed. We initially scored this 3 of 3 and corrected it the same day: Perplexity’s pricing answer had been counted as retrieval when every fact in it was available off-site.

Week of 30 June – 6 July 2026 (Month 2 audit — fired on schedule, no score movement)

A few days late this month — I was off-grid with patchy internet. The July 4 audit fired on schedule; the write-up only landed today. Discipline resumes.

• Monthly Citation Audit fired on schedule Saturday 4 July at 10:30am AEST. Result: 0/12 across ChatGPT, Claude, Perplexity, and Google AI Overviews. Same as Month 1. Same as Day-0.
• Discovered the Monthly Refresh workflow — the automation that pushes results to this page — is producing hallucinated narrative and impossible scores (an “18/12” and false “citations tripled” text). Manual publish only until fixed. Debug queued for the week before the August 1 audit.
• Off-grid for a fortnight — no on-site citation-targeted content deployed this cycle. The audit ran itself; the shipping didn’t.

Observation: Month 2 zero. That’s an honest signal — not that AI engines are ignoring us, but that we haven’t given them anything new to cite in six weeks. The infrastructure work of May–June (thank-you page, privacy policy, ToS) went off-site and doesn’t move a citation score. On-site content ships this fortnight or the August 1 audit reports zero for the third month running. The next four weeks are the test.

Week of 30 May – 6 June 2026 (Month 1 audit — delayed 3 days, no score movement)

• Ran the Month 1 catch-up audit on 9 June (target was 6 June). Result: 0/12 across all four engines — no change from Day-0
• Built the SAM thank-you page with first-name personalisation and conversion pixel firing (Meta Lead, GA4 generate_lead, LinkedIn conversion). Required before Phase I ad launch
• Published the privacy policy and terms of service drafts — required for Meta/LinkedIn ads approval and AU Privacy Act compliance
• Identified the gap that caused the 3-day audit delay: no Monthly Citation Audit workflow exists yet. Building it this week so 4 July fires on timeCaptured the official Day-0 baseline via the Phase F workflow: 0/12 citations across ChatGPT, Claude, Perplexity, and Google AI Overviews on the three locked buyer queries

Observation: shipping the infrastructure was the unsexy work of the past three weeks. No new on-site citation-targeted changes deployed this week — the schema typo fix, H1 add, and 6 FAQ blocks shipped in the 5–9 May window (see entry below). The June 6 audit will be the first true test of whether those May fixes registered with the engines in 4 weeks. If they did, expect 1–3 citation lifts. If they didn’t, the next 4 weeks of weekly schema work will compound by July.

Week of 26 May 2026 (delayed page launch — Day-0 officially captured)

• Captured the official Day-0 baseline via the Phase F workflow: 0/12 citations across ChatGPT, Claude, Perplexity, and Google AI Overviews on the three locked buyer queries
• Locked the three priority queries permanently — “AI marketing consultancy for professional services Sydney” / “AI search visibility audit for small business” / “How do I check if my business shows up in ChatGPT”. These won’t change for the life of this log
• Deployed this page — the public log starts here. Three weeks delayed from the original 5 May plan because the audit infrastructure (Railway citation service, Snapshot delivery pipeline, monthly refresh workflow) took longer to wire than expected
• Built the production-grade SAM Citation Service on Railway — same pipeline that runs every client Snapshot now runs the monthly self-audit

Observation: shipping the infrastructure was the unsexy work of the past three weeks. No new on-site citation-targeted changes deployed this week — the schema typo fix, H1 add, and 6 FAQ blocks shipped in the 5–9 May window (see entry below). The June 6 audit will be the first true test of whether those May fixes registered with the engines in 4 weeks. If they did, expect 1–3 citation lifts. If they didn’t, the next 4 weeks of weekly schema work will compound by July.

Week of 5–9 May 2026 (early schema fixes — pre-launch foundations)

• Rebuilt the homepage Schema.org block: added LocalBusiness with NAP + Sydney + Australia areaServed, Person schema for Rowan Shead, FAQPage with 6 Q-and-A pairs, fixed the Organization name typo “Strategi Ai Marketing” → “Strategic AI Marketing”
• Added a single H1 to the homepage: “AI Marketing Consultancy for Professional Services in Sydney”
• Deployed 6 visible FAQ blocks on the homepage matching the FAQPage JSON-LD — answer capsules for the buyer queries our prospects type into AI engines

Observation: the schema typo fix alone is expected to surface in Perplexity within 2 weeks based on prior client data. This week shipped the structural foundations — the work that has to happen before any monthly chart can move. The three buyer queries used in these FAQ blocks were later refined and locked permanently on 26 May during Phase F; the FAQ blocks were updated then to match. The 6 June audit will be the first measurement of whether these foundations register with the engines in the first 4-week window.

The Current Experiment

The Current Experiment Each month we pick one variable to test on strategicai.marketing itself before we recommend it to clients.

Month 5 — audit Saturday 7 November 2026

Hypothesis: the three Track B cells are the owner pages working, not one engine’s variance. If that is right, Perplexity holds at least two of the three on 7 November and at least one other engine cites the site on any of the six queries.

Method: nothing changes about the measurement. The work is three overdue plays: (1) the how-to post for “how do I check if my business shows up in ChatGPT”, three months late; (2) deploy and validate the schema already written for the homepage, How This Works and the Visibility Index; (3) move How This Works inside /marketing-for-tradies/ with a redirect, so query 2’s owner page sits in the cluster that supports it.

Control: Google has been able to read the site the entire time and has never cited it on either track.

On the record: if Track B falls back to zero or one in November, the three was variance and we will say so. If it holds and no second engine moves, it is a Perplexity result and not yet a content result, and we will say that instead.

And one thing we are putting on the record now, before we know the answer. Three months of zero have had an explanation attached each time — infrastructure work that went off-site, a fortnight off-grid, a firewall. All three were true. If September is zero, there is no explanation left, and we will publish it as the content not working.

What To Do Each Week To Lift Your Own Score

Same playbook we run on ourselves. Five hours a week. Twelve weeks to a meaningful score lift.

Monday — 30 min — Audit drift check

Run a single buyer-intent query on each of the four AI engines. Log the result (date, engine, query, cited Y/N, who got cited if not you). Twenty minutes of work that catches sudden drops.

Tuesday — 90 min — Ship one schema fix

Pick the highest-strategic-value page. Add or upgrade ONE schema type: Article, FAQPage, Service, LocalBusiness, Person, or BreadcrumbList. Validate via Schema.org and Google’s Rich Results Test before deploying. One page per week. By week 12, your 12 most strategic pages are upgraded.

Wednesday — 60 min — Write one answer capsule

Pick one buyer question. Write a 40–80-word self-contained answer. Place it under the relevant H2/H3 on the most relevant page. Style as a blockquote or pull quote. One per week. By week 12, you have 12 capsules — the citation surface area most small-business sites lack entirely.

Thursday — 60 min — Topical depth move

Pick one priority topic. Add 300–500 words of depth to your STRONGEST existing page on that topic. Depth, not breadth. AI engines reward pages that go deep, not sites that go shallow on many topics.

Saturday — 30 min — Weekly log + one observation

Update your own we-did-this log: date, 1–3 bullets of work shipped, one observation. Future-you will thank present-you for the dated record.

Total: 4.5 hours per week. Twelve weeks. ~54 hours of focused work.

How Do I Improve AI Search Visibility?

To improve AI search visibility, follow this 4.5-hour weekly playbook for 12 weeks. Monday — query each of the 4 AI engines with one buyer question and log results (30 min). Tuesday — ship one schema type to your highest-priority page (90 min). Wednesday — write one 40–80-word answer capsule (60 min). Thursday — add 300–500 words of topical depth to your strongest page on a priority topic (60 min).

Saturday — log what you shipped and one observation (30 min). Expected score lift over 12 weeks: 3–6 citations gained out of 12. This is the exact playbook Strategic AI Marketing runs on itself, publicly logged at strategicai.marketing/we-did-this.

How The Audit Runs

The same audit pipeline runs on strategicai.marketing that runs on every Snapshot delivered to a client. No special tooling. No analyst tweaks. No outlier removal.

The four engines: ChatGPT (OpenAI Responses API with web search), Claude (Anthropic Messages API with web search tool), Perplexity (Sonar API, search-grounded), Google AI Overviews (Serper.dev, AU geo + en).

The three locked buyer queries (locked 26 May 2026, never changed):

1. AI marketing consultancy for professional services Sydney — commercial intent
2. AI search visibility audit for small business — category / authority intent
3. How do I check if my business shows up in ChatGPT — diagnostic / pain-point intent

The scoring rule: a citation counts if the engine names Strategic AI Marketing in the answer text OR includes a direct link to strategicai.marketing in the cited sources. Implicit references (“an Australian consultancy”) don’t count.

The audit day: first Saturday of each month, 11am AEST. Same time every month removes time-of-day variance from results. First monthly re-audit: Saturday 6 June 2026.

Honest disclaimers:

• AI engines change retrieval algorithms without notice. A score drop month-to-month doesn’t always mean we did worse — sometimes the engine changed.
• The 12 buyer queries we chose are our category. They’re not yours. Your Snapshot uses 3 buyer queries from your category.
• Citations on a single Saturday morning aren’t perfectly representative. They’re a snapshot — same as everything else here.

How do I check if my business shows up in ChatGPT?

To check if your business shows up in ChatGPT, open ChatGPT with web search enabled and type a buyer-intent question your customers would actually ask — for example, “best [your category] in [your city]” or “how do I find a [your service].” Watch whether your business is named in the answer text or linked in the source list. Repeat across Claude, Perplexity, and Google AI Overviews.

A citation only counts if the engine names you directly or links your domain — implicit references don’t count. Strategic AI Marketing’s free Snapshot runs this check across 4 engines × 3 buyer queries automatically and delivers results in 60 minutes.

Why This Page Exists

Three reasons:

1. An honest record. Nobody else in our category publishes their own AI-citation score. We’re publishing ours because we’d rather be the consultancy that did, even when the early numbers were embarrassing.
2. A playbook other people can run. The weekly action plan above is the same one we use. Take it. Run it on your own site. Your version of this page can exist by Christmas.
3. Proof for the Snapshot landing page. When we say “we use this on ourselves first,” this is the page that proves it. The score, the dates, the work shipped — all verifiable.

The Snapshot you can request from the main landing page is the same diagnostic that produced every score on this page.

Get Your Free Snaphot

• Last monthly audit: 3 October 2026 (Month 5 — Track A 1/12 · Track B 3/12)
• Last log entry: Weeks of 6 September – 3 October 2026
• Next monthly audit scheduled: Saturday 7 November 2026

If you’re reading this more than 6 weeks after the “last weekly log entry” date, something’s gone wrong with our cadence. Email Rowan and tell us — we’d rather know.

---

Frequently Asked Questions

© 2026 Strategic AI Marketing

---