AI built this site. Here's the live ranking data.

Repo initialised. Brief committed as the north star. Astro + Tailwind scaffold, full launch page set and first blog posts built by AI in a single session. Clock started.

Site deployed to production and live at ai4smb.com.au - same session as repo-init. The AI verified DNS, SSH and server config (nginx on a Virtualmin VPS), diagnosed a trailing-slash redirect clash with the canonical URLs, and shipped the fix. Repo-init to live: under 2 hours.

Day 1: 13 AI-generated illustrations placed site-wide, and Phase 2 shipped - vertical pages for tradies and clinics, plus Melbourne, Sydney and Brisbane location pages. Google Search Console verified on day 0; sitemap processed the same day. First backlink live (the founder’s announcement on pearceit.com.au).

CI pipeline live: pushes to main now build and deploy automatically. Privacy-first analytics online - the AI installed self-hosted Umami on a fresh Linux container (diagnosing a package-manager version clash along the way), wired it through the reverse proxy, and verified the first pageview end-to-end. No Google trackers, no cookies, no consent banner needed.

Day 7: first Search Console data. 161 impressions across 49 queries in the first five days - impressions started on launch day itself. "ai receptionist australia" (the flagship target) is on the board at average position 83; the branded term is already at 15. Zero clicks yet, exactly as expected at these positions. The tracker below now shows Search Console 7-day average positions as the honest baseline; dedicated weekly rank tracking takes over from here. Also fixed this week: the HTTP version of the site was indexed separately - now 301s to HTTPS.

Day 8: the research payload landed. The AI transcribed and analysed 1,972 short-form videos (~500k words) from two SEO/GEO practitioners, cross-referenced their claims into a playbook - where both independently agree became site changes, where they conflict became the experiments below. Same day: every page and post retrofitted to answer-first shape, and an AI-citation tracker went live - 42 prompts an Australian SMB owner would actually type, run against ChatGPT with search, logging which sites get cited. The day-8 baseline is honest and brutal: ai4smb.com.au cited zero times, mentioned zero times; even our brand prompts resolve to a German consultancy with the same name. Reddit is the single most-cited domain. Now we know exactly what the mountain looks like.

The citation data went straight back into the site: a dedicated 3CX integration page (the one prompt surface no vendor owned), a dental vertical page, and a plain-English guide to the law on AI answering calls in Australia - a question currently answered only by government sources. Then the AI ran its first fully autonomous overnight session on a standing grant from the founder: build, merge, deploy, verify - with the human reviewing the diffs over breakfast. This entry was written during that session.

Measurement change, published because the method matters as much as the numbers. AI answer sets get rebuilt constantly - a third-party study doing the rounds puts the median citation’s life at 11-15 days, with 44% of sources cited exactly once and never again. We have not verified that study independently, but the churn it describes matches what our own weekly runs show. Two consequences. A reading taken only at day 30 is close to a coin flip for any single prompt, so the AI arm of E1 is now read every fortnight. And “were we cited at all” is the wrong number, so we now report share of visibility - of all the times an AI answers a question in our space, how often the cited source is us rather than a competitor. The day-90 target moved with it: it now requires the result to hold across two consecutive readings, not one lucky one.

Went after the day-30 entity target. At the day-8 baseline, asking an AI engine “What is AI4SMB?” returned a German software boutique and a Texas firm of the same name - never us - so the first job is not visibility but identity. Shipped: the Organization schema on every page now carries the ABN, the public Australian Business Register record and the PearceIT link as sameAs, the registered legal name, a Victorian address and an explicit disambiguatingDescription; /about opens with an answer-first “What is AI4SMB?” block naming the businesses we are not; a four-question identity FAQ with matching structured data; and llms.txt now leads with the same statement. Every surface an engine reads now says the same thing. Then re-ran the full 42-prompt set the same afternoon. We are now cited in 5 of 42, against 0 of 42 at the day-8 baseline. All three brand prompts land: ai4smb.com.au is described first, correctly, as the Australian agency founded by Matt Pearce, where at baseline the answer was the German and Texan businesses only. The engines still list those alongside us, which for a genuinely shared name is the right answer rather than a failure. The other two hits are “AI consultant Sydney” and “AI consultant Brisbane” - and pointedly not Melbourne, the city we actually work from, which is a useful reminder that these answers are not built the way a local directory is. Everything else - money, cost, comparison, vertical, informational, objection - is still zero. So: the identity problem has moved, the commercial problem has not, and the day-60 target of a first non-branded money-prompt citation is a real target rather than a formality. One caveat we will not paper over: this was the first re-read since day 8, so it measures nineteen days of indexing plus that morning’s entity work, not the entity work on its own. We cannot claim the schema did it.

The afternoon’s citation probes went straight back into strategy, same day. The money-prompt map showed the PBX-integration surface is the open door (the 3CX prompt has no third-party incumbent) while the AU-accent surface is locked four competitors deep - so the differentiator work shifts from the voice alone to the integration layer. Shipped accordingly: a dedicated Yeastar page, written from inside our founder’s own live 3CX-to-Yeastar migration - Yeastar’s native AI ships Australian voices (two male, two female at the time of writing) and has begun exposing MCP, which makes it the first PBX where an AI agent can act on the phone system rather than just answer it. Yeastar is now co-primary across the site; 3CX remains fully supported, and its page now discloses the Yeastar voice difference because buyers deserve the comparison. Also planted: a plain-English definition of “managed intelligence” - a term at roughly ten searches a month in Australia, which is the point: the tracker gained a prompt for it (and one for Yeastar), so we will watch both surfaces form from birth.

First public pricing, hours after the competitor analysis that demanded it. Studying the one agency that seated itself in an AI answer core within months of founding showed transparent pricing doing citation work our “pricing being finalised” could not. The founder set the first two numbers from a live client quote the same afternoon: the AI Opportunity Audit - audit, analysis and written report - from $800 ex GST, and hosted Yeastar systems from $150/month ex GST. Both are now on the site in text and in structured data (a real minPrice specification, GST flagged). The AI receptionist’s monthly rate stays honestly unset until it is set with the founding cohort - a discount for being the proving ground is only worth something if the proving is real.

Repositioned, on the founder’s call and the day’s data. The site launched receptionist-first; the citation probes showed the receptionist prompt surfaces are the most locked (the Australian-accent prompt alone has four incumbents in its stable core) while the category surfaces - best AI automation agency, who can help my small business start with AI, managed intelligence - are the open ones. So the hierarchy inverted: the brand frame is now the category (an AI automation agency run with MSP discipline), and the AI Receptionist becomes flagship product #1 under it rather than the site’s identity. Hero, tagline, schema slogan and OG copy updated; every receptionist page, post and tracked keyword stays exactly where it was, because the product is not being demoted - the roof over it is being widened. The tracker measures whether the engines agree.

Day 50, and the first entry that is mostly an admission. The day-30 checkpoint was due on 8 August and is being written on 28 August. In between, the site kept building and deploying cleanly, but nobody measured anything: the citation tracker has two full reads on record, 17 July and 5 August, against a doctrine adopted on 5 August that says read it fortnightly. Publishing cadence decayed the same way - six posts on 9 July, four on 18 July, two on 5 August, one on 20 August. The thesis being tested here is that an AI can build *and maintain* a commercial site. Building is going fine. Maintenance is the part that slipped, and since the whole point of this page is to publish the result either way, it goes on the board. Worse, the rank table on this page was still showing 16 July numbers, which made us look better than we were: it listed “ai receptionist australia” at 83 when the term is not in the top 100, and the branded term at 15 when it is at 6. One of those errors flattered us and one did not, which is exactly why undated tables are dangerous. The table above now carries the date it was pulled and the previous reading beside the current one, so the direction of travel is visible instead of implied.

Added Google Analytics 4 alongside Umami, which is a deliberate walk-back of something this log said on 13 July. That entry celebrated “no Google trackers, no cookies, no consent banner needed”, and it was true when it was written. It is not true any more, and the honest fix is to say so here rather than quietly edit the old entry. The reason for the change: on 13 May 2026 Google added a native “AI Assistant” channel to GA4, which classifies ChatGPT, Gemini and Claude referrals as their own traffic channel. This site is an experiment in AI-surface visibility, so that is close to the single most useful measurement available to it, and GA4 processes forward-only - every day without the tag is AI-referral data that cannot be backfilled. Umami stays as the privacy-first primary and already showed the raw referrals (six visits from chatgpt.com and one from claude.ai over the last 60 days); what GA4 adds is the channel breakdown, not the underlying visibility. Three limits worth stating plainly, because the number will be wrong in a specific direction: roughly 60 to 70% of AI-assistant sessions arrive carrying no referrer at all and are counted as Direct, Google AI Overviews are counted as Organic Search rather than AI Assistant, and Perplexity lands in plain Referral. GA4 undercounts AI traffic too. The privacy page now names both tools, says which one sets cookies, and links Google’s opt-out.

Then the measurement itself turned out to be pointed at the wrong thing, and fixing it produced the best result of the experiment so far. The prompt set was built at day 8, when this site was receptionist-first. On 5 August the business repositioned category-first, with the receptionist as flagship product #1 underneath. The prompt set never followed: 28 of 44 prompts were phone-shaped, and not one asked about what we would actually do for a business beyond the phone. No audit prompts, no workflow automation, no self-hosted, no managed services, no staff adoption. We were still measuring the old identity, and had we reported on it we would have concluded the repositioning failed without ever having tested it. Added 17 category-level prompts covering the real offer ladder. No existing prompt was edited, only new ids, so every series back to the day-8 baseline still joins. Then ran the full 61 against ChatGPT with web search. Result: cited in 11 of 61, every one verified as a genuine ai4smb.com.au citation rather than a name match. On the 44 prompts that also existed on 5 August, it is 9 against 5 a fortnight ago. The split is the finding. Category prompts: cited in 6 of 30. Receptionist prompts: 2 of 28. Brand prompts: 3 of 3. Category visibility is running at roughly three times receptionist visibility, which is exactly what the 5 August citation core-map predicted and precisely what the old prompt set could not have shown us. Four surfaces we were absent from on 5 August now cite us: "best AI automation agency in Australia for small business", which the core-map had flagged as the most open money surface in the field; "AI receptionist that integrates with 3CX", the PBX surface no vendor owned, where the dedicated page did its job; the Cliniko integration prompt; and "what is managed intelligence", which is the one we care about most. That last one was a bet placed on 5 August: publish a definitional page for a term with roughly ten searches a month in Australia and see whether we can seat ourselves in a surface as it forms. Twenty-three days later, we are cited in it. Two of the seventeen brand-new prompts cited us on their very first run, both from the managed and automation side: "who can automate my business admin with AI in Australia" and "managed AI services for Australian small business". Nothing that cited us on 5 August stopped. And the field position has moved in a way we did not expect this early: across all 61 answers, ai4smb.com.au is now the equal second most-cited domain with 11 appearances, level with reddit.com and behind only techradar.com on 17. At the day-8 baseline Reddit was the single most-cited domain in the space and we appeared zero times. The pages doing the work are the homepage, cited seven times, and /services, cited four.

The unflattering half of the same read, published alongside it per the rule. Fifteen of the seventeen new category prompts returned nothing at all, and the pattern in which fifteen is uncomfortable: the entire audit front door is invisible. "I want to use AI in my business but don’t know where to start", "AI opportunity assessment for an Australian small business" and the readiness-assessment cost prompt all return other people. That is the top of our offer ladder and the destination of the site’s primary call to action, so being absent there costs more than being absent from a vertical page. Self-hosted AI is invisible too, despite being one of the four stated differentiators and the reason a clinic or a law firm would pick us. Melbourne remains absent while Sydney and Brisbane both cite us, unchanged since 5 August and still the oddest finding in the dataset, since Melbourne is where the founder actually works. On the Google side nothing has improved: fifteen of eighteen tracked terms are outside the top 100 and the flagship term is not ranking at all, so the day-60 ranking target will be missed. And the honest caveat on the good news: this is one reading. Our own doctrine says a result must hold across two consecutive readings before it is called, precisely because AI answer sets churn, so the correct description of 11 of 61 is a strong single data point and not yet a trend. The seventeen new prompts have no baseline at all, so their first-run numbers are a starting line rather than a movement. The next read is 15 September and it is now automated rather than dependent on anyone remembering.

Last thing, and it is the one that surprised us. Before switching the tracker to a newer model we ran a calibration: the same eighteen money prompts, the same afternoon, asked three times through three different model tiers. The assumption being tested was the obvious one, that a cheaper model would search less thoroughly and therefore under-report our visibility. That is not what happened. The two-year-old gpt-4o returned an average of 9.0 cited domains per answer and 113 unique domains across the eighteen. The current mid-tier model returned 2.7 per answer and 29 unique. The current budget model returned 4.7 per answer and 60 unique. The newest mid-tier model is the one that cites least, by a wide margin, and it is not close. Which means the choice of measuring instrument moves the headline number more than a fortnight of content work does: on the identical prompts, on the identical day, against the identical website, we appear in 4 of 18, 2 of 18 or 5 of 18 depending purely on which model is asked. A caveat we will not skip: at one sample per prompt those citation counts are not distinguishable from the churn we already documented, where roughly half of any cited list rotates between asks. The retrieval-depth gap is a different matter - a three-fold difference in domains per answer, consistent across the whole set, is structural rather than noise. So the decision rests on that. We are staying on gpt-4o, which is deliberately not the newest option: it preserves the series back to the day-8 baseline, and it surfaces three times as much of the competitive field per question asked, which is the part of this dataset that actually tells us where to point the next piece of content. A realism check against the current flagship gets run quarterly, because an old model is a fine yardstick right up until it stops resembling what a buyer is served.

Two housekeeping corrections worth publishing because both were errors in our own favour. First, collection is now a scheduled job that commits its raw dataset to the repository on the 1st and 15th, whether or not a human or an AI remembers. The day-30 miss happened because measurement depended on someone choosing to run a script, and the lost weeks cannot be backfilled. Interpretation stays a separate judgement task; the collection half should never have needed one. Second, the running cost of this experiment was being quoted internally at about a dollar a run and fifty dollars a year. That figure came from an estimate written in the tracker’s own README and never checked against a bill. The actual billing export shows the 5 August session cost USD $5.91 for 93 requests, which is 6.35 cents a prompt, dominated not by the answer but by the roughly 17,500 tokens of retrieved web content fed into every single request. At 61 prompts that is $3.87 a run and $201 a year weekly. The cadence is therefore now fortnightly, which is what our own published measurement doctrine specified in the first place, halving that to about $101. A number nobody had verified had been repeated confidently twice.