Experiment 001: our own site, read in the open.

The question

Every page on this site says a business has to be understood twice: by the machines that decide who gets found, and by the buyer who arrives. If that is true, it has to show on our own site first. Before the rebuild, Google showed seoinformatica.com mostly for two guides about Google’s AI answers and for a pricing article, and almost never for the work we sell. The question is whether a site rebuilt on the method changes that, and by how much.

The hypothesis, as written on 2 October 2026

Built on the method, this site is found for AI SEO and the searches around it, not only for its name, and the AI assistants name it when someone asks who to use. United States first.

The prediction, in the terms our method uses for every change: the rebuilt pages are refreshed pages, so they predict more clicks; the pages added in October are new pages, so they predict more impressions. The claim is about searches that are not our name. A rebuild that only makes the brand easier to find has not done its job.

The change

Date (2026)	What changed
29 September	Baseline recorded. The positioning, the offers and the content direction reset.
30 September	The site moved from WordPress to our own site kit on Cloudflare Pages. Every address kept or redirected.
2 October	Phase two: the new positioning on every page, one job per page, every claim with its source.
3 October	The home page and ten core pages redesigned as paper exhibits; share cards; llms.txt and llms-full.txt.
4 and 5 October	Titles and headlines of 42 pages rewritten for B2B service businesses; the company described by how it works, not by who runs it.
6 to 9 October	36 pages added: nine AI SEO guides, twenty industry guides, five diagnostic guides, two website resources. A site search. The publication dates of 39 earlier pages set to their October update dates.
10 October	Enquiry forms protected, with each enquiry’s source recorded.
13 October	The free AI Crawler Checker, with links from four guides. The last change before the reading.

Everything after 2 October is part of the same experiment. We did not hold the site still, so the reading is of the whole change, not of any one piece of it.

The baseline: Google, 4 September to 1 October 2026

The 28 days before the rebuilt site went live, from Search Console (web search, final data, the whole property). The file linked at the top holds every row.

Measure	Clicks	Impressions	Average position
Whole site	12	7,000	14.1
Searches containing our name	2	17	8.9
Searches that are not our name (visible queries)	1	2,531	27.1
United States, not our name	0	1,379	30.1

Google hides rare queries, so the two filtered rows cover the 286 queries it shows; 9 clicks and 4,452 impressions sit in queries it does not name. Eight of the 12 clicks came from India, most of them our own name; the United States gave 3,602 impressions and no click, the United Kingdom 540 and none.

Where the impressions were: /google-ai-overviews-what-matters/ (2,506), /ai-visibility-cost/ (1,160, 2 clicks), /track-chatgpt-referrals-ga4/ (791, 1 click), /google-ai-overviews-myths/ (658), /technical-seo/ (639, average position 46.6). The pages that sell the work had little: /local-seo/ 147 impressions at position 35.8, /website-rebuild/ 134 at 36.4, /new-website/ 31, /start/ 16, /ai-seo/ 4. The Outside View page did not exist yet. Together the eight decision pages had 1,218 impressions, 2 clicks.

The largest searches that are not our name: “seo funnels” (186 impressions), “ai visibility” (121), “alternatives to technical seo” (99), “ducr” (98), “ai cost visibility” (92), “seo for technical support software” (85), “export technical seo” (77), “how to track chatgpt traffic in ga4” (65), “website rebuild” (58) and “ai visibility pricing” (48, the one click, position 3.8). Several of the longest queries in the file are prompt-shaped strings sent through Google by AI tools, not by people; they add impressions and no clicks.

The baseline is not flat. Impressions ran at 300 to 443 a day from 5 to 16 September and at 106 to 203 a day from 18 September onward, before anything changed. The daily series is in the file.

The baseline: the assistants

We did not record what ChatGPT and Claude said about us before the rebuild. That is a gap, and we say so rather than reconstruct it. The first recorded run, due with the first look on 27 October, is an observation, not a baseline; a change in what the assistants say is claimed only between two recorded runs after that.

The protocol: six questions a buyer would ask, in the words a buyer would use (which agency to hire for AI SEO for a consultancy in the United States; who to ask for a service website that gets found; what an Outside View or a free audit of a service site is worth; and three more fixed in the log), each asked three times in ChatGPT with search on and three times in Claude with web search, on one day, from a fresh conversation. We record the date, the model, whether SEO Informatica is named, where in the answer, and which pages are cited. The log is kept in the same format as our citation tracking log.

What we watch, and the decision rules

Written on 13 October 2026, before any reading. The reading is judged against this text, not against what the site says then.

1. Searches that are not our name. Clicks and impressions for the whole property with the query filter “does not contain informatica”, the same filter as the baseline. Confirmed needs clicks above the baseline’s 1 and impressions above 2,531 in the 28-day window, with the United States above its 0 clicks and 1,379 impressions.
2. The decision pages. Impressions and clicks on the eight pages that sell or start the work: /ai-seo/, /local-seo/, /technical-seo/, /content-funnels/, /website-rebuild/, /new-website/, /outside-view/ and /start/. Confirmed needs their combined impressions above 1,218 and their impression-weighted average position better than the baseline’s.
3. The assistants. Whether ChatGPT or Claude name SEO Informatica for at least one of the six questions in a recorded run. Reported alongside the verdict, not part of it until two recorded runs exist.

The verdict follows our standing rule. Confirmed: rules 1 and 2 both move the predicted way, with comparable evidence. Refuted: comparable evidence and both move the other way. Unclear: one moves and one does not, no movement, or evidence that is not comparable. Zero movement reads unclear, never confirmed. A result that comes from the hidden queries alone, which we cannot filter, is reported but does not confirm rule 1.

When we read

The site kept changing until 13 October, so the clean settling period runs to 27 October. The first look on that day is a check that the instrument works and that nothing broke; it is not a verdict. The comparison window is 16 October to 12 November, the 28 days that follow the settling period written down on 2 October, read when Search Console’s final data covers it; the verdict is published with the reading on 13 November. If the window turns out to include a change we did not plan, the reading says so, and a second window, 27 October to 23 November, is read as well.

Readings

Reading	Date	Result
First look	27 October 2026	Pending.
28-day comparison and verdict	13 November 2026	Pending.

What this reading cannot tell you

• Many things changed at once. The reading is of the whole rebuild and everything shipped behind it, not of any single change, and it cannot say which part did what.
• Twenty-eight days is a short window. It is directional, not seasonal truth. September already showed a fall in impressions before the rebuild, and we do not know why.
• Thirty-six pages were added. More pages predict more impressions almost by construction; that is why rule 1 is about searches that are not our name and rule 2 is about the pages that sell the work, not about totals.
• Google hides most of the queries behind our clicks. Any change in the hidden share could move the filtered figures without the search behaviour changing.
• Some impressions are not people. Prompt-shaped queries from AI tools sit in the baseline and will sit in the reading.
• The assistants were not recorded before the rebuild. Their first run is an observation.
• This is one site, our own, and we built it. A confirmed verdict says the method worked here, once; it is not a promise about anyone else’s site.

Check it yourself

The baseline file holds the totals, the brand split, every page, the twenty largest non-brand queries, the countries and the daily series, as pulled on 13 October 2026. Anyone with Search Console can repeat the pull for their own site with the same settings: web search, final data, a 28-day window, the property total, then the query filter “does not contain” your own name. Our readings will be published the same way, with the file.