Zero to Cited: a public evidence log

Objective and start date

Run Robot Visible's own AI visibility as a measured experiment, in public, on a question panel that does not change while the programme runs. The point is not to publish a good result. It is to publish a comparable one, including the weeks where nothing moves, so that anyone reading it can tell the difference between evidence and a claim.

The programme started on 17 August 2026. This page was published on 16 August 2026, the day before the first questions were scheduled. The first production answer check appears below. It is a dashboard-level baseline; the stricter weekly record remains empty until its answer excerpts, per-answer timestamps and source captures can be published in full.

The fixed question panel

Twenty non-branded questions, frozen from 17 August 2026 and unchanged for the length of the programme. A panel that can be edited mid-run is a panel whose results can be improved without improving the site, so it is not edited.

Question-to-page ownership

Every non-branded question has one primary page, assigned before the first run. Other pages may link to the owner. They do not compete with it, and a question without a credible owner is a reason to improve the mapped page rather than to publish another URL.

The branded controls map to the homepage, /pricing and /alternatives/peec-ai as appropriate. They are listed here for completeness and are not part of the non-branded ownership map above.

The surface matters as much as the provider, because an API answer and a consumer app answer are different measurements and only one of them is being taken here.

This list is the provider and surface set the initial production answer check used. It is read from the dated dashboard baseline below. Four Google Gemini answers were unreadable and are shown as unavailable, not absent.

A provider that errors, rate-limits, or is not checked in a given week is recorded as unavailable for that week. Unavailable is not a negative result, it is not counted as a miss, and it never enters a denominator.

Baseline, 16 August 2026

Supporting AI citation evidence from Bing AI Performance, reported with what it proves and what it does not.

Bing's report is a platform aggregate for Microsoft AI experiences and carries no query dimension. It is supporting AI citation evidence, not a query-level panel result, a claim about ChatGPT, or evidence about a customer.

Initial answer check, 18 August 2026

The useful result is not the dashboard's combined percentage. Robot Visible appeared in 0 of 72 readable non-branded answers. Eleven questions had at least one monitored competitor winning and eight were open with no monitored brand winning. Those are reported as 11 competitor-winning questions and 8 open questions, not blended with the branded controls.

The four branded controls returned 16 readable answers: 15 citations and one mention, with no recommendation recorded. They show that the providers could retrieve and describe Robot Visible when the name was supplied. They do not show non-branded discovery.

Across all 23 questions, 88 of 92 provider answers were readable. Four Google Gemini answers were unreadable and are excluded rather than counted as absent. Question 17, “how do indie hackers get their product mentioned by ChatGPT”, was not added to this production run and has no result. It will not be backfilled into this baseline.

The 11 questions where competitors supplied the cited pages

These are the highest-value comparisons in the first run. The URLs below are representative competitor pages preserved in the expanded provider evidence, not an assertion that every provider used every URL.

Run timestamp

18 August 2026 at 17:35, as displayed by Robot Visible; the dashboard capture does not print a timezone.

Captured evidence

Three Robot Visible dashboard captures plus the saved expandable question evidence. robotvisible.com; latest completed answer check; OpenAI, Claude, Perplexity and Google Gemini; checked 18 August 2026.

No public capture yet. The aggregate and competitor URLs can be checked inside Robot Visible, but the screenshots and full answer excerpts have not yet been published for outside review.

Weekly results

Three separate states, reported in three separate sections. None of them is evidence for another, and none of them is combined into a score. Where a state has not been measured yet, it says so.

Change ledger

Every material change to a monitored page, with its date. A change and a later movement are not connected by their order, so this exists to make coincidences visible rather than to claim causes.

1. 18 August 2026

   Zero to Cited: a public evidence logWhy ChatGPT cannot find your website

   Published the first production AI answer check as a dashboard-level baseline: 23 questions across OpenAI, Claude, Perplexity and Google Gemini, with branded controls separated from non-branded discovery, four unreadable Gemini answers excluded rather than counted as absent, and the 11 competitor-winning questions linked to representative cited pages. Question 17 was not in the production run and is recorded as not measured. The full answer excerpts, per-answer timestamps and public dashboard captures are not yet published, so this entry does not claim to be a complete weekly Observed record. Updated the discovery guide's link to describe the initial result rather than saying the evidence log still held a baseline only.

   The first live answer check completed on 18 August, and publishing its null non-branded result is the programme's first required evidence update.

2. 16 August 2026

   Why ChatGPT cannot find your website

   Added direct answers to the five questions the page owns, a stage-by-stage decision table covering access, indexation, understanding, selection and lost citations, two constructed worked examples, first-party crawler sources, and contextual links to the crawler reference, the Cloudflare, ChatGPT, Perplexity and Google guides, and the free scan. The URL, canonical and primary intent are unchanged. The examples are written illustrations on a fictional domain, not redacted captures, and the plan's request for original redacted examples is still outstanding.

   The page is the assigned owner of panel questions 9, 10 and 12, and those questions were not answered directly anywhere on it.

3. 16 August 2026

   Robot Visible vs Peec AI

   Replaced the analysis-versus-delivery framing with a precise account of what each product analyses, added a page-level comparison method with its limits, added explicit better-choice lists for both products, and linked on to the tool-choosing guide and Robot Visible pricing. No Peec AI price, allowance, model count or product claim was changed, and the pricing verification date is unchanged. The refresh is incomplete: Peec's prices render client-side and could not be re-read at source, so the price re-verification the plan requires has not succeeded and the page says so.

   The page is the assigned owner of panel question 8, and its old wording understated the analysis Robot Visible actually performs, which made the comparison inaccurate against ourselves.

4. 16 August 2026

   Zero to Cited: a public evidence log

   Published this evidence log with the objective, the frozen question panel, the ownership map, the providers the panel will be run against, and the 16 August AI citation baseline. It ships with the baseline only, no run has taken place, and the baseline capture is not yet published.

   A public programme that reports on itself needs a fixed, dated place to be wrong in.

Methodology

1. 1.Freeze the panel. The 24 questions are fixed from 17 August 2026 and are not edited while the programme runs, because a panel that changes cannot produce comparable weeks.
2. 2.Give every non-branded question one primary page, recorded before the first run. Supporting pages link to the owner instead of competing with it.
3. 3.Run the panel through every configured production provider on the same day each week, and record the provider, surface, model where one applies, and timestamp for each result.
4. 4.Record what actually came back: the answer, which brands were mentioned, which were recommended, which were cited, and every source URL the answer used.
5. 5.Keep Eligible, Competitive and Observed in three separate records. No stage is ever used as evidence for another.
6. 6.Record provider errors and unchecked providers as unavailable. Unavailable never becomes a negative result and never enters a denominator.
7. 7.Export Bing AI Performance for the same weekly window and keep it separate from the question-panel results as supporting AI citation evidence.
8. 8.Log every material page change here with its date, so a later movement can be read against what was shipped instead of attributed to it.

Limitations

• This is one site measuring itself. It is a worked example of a method, not a study, and nothing here generalises to a category.
• Answers vary by provider, surface, model, time, location and personalisation. One answer at one moment is a sample.
• API surfaces are not consumer apps. A result from a provider's API does not describe what a person using that provider's own product sees.
• The Bing figure is a platform aggregate with no query dimension. It cannot be turned into a query-level result, a brand mention, a recommendation, a ChatGPT result, or a customer outcome.
• The baseline capture is not published yet, so the baseline is currently Robot Visible reporting its own dashboard. Until the export is here, treat its figures as unchecked by anybody else.
• A page change and a later movement are not causally linked by their order. Where the ledger and a result line up, that is a coincidence worth investigating, not an attribution.
• Publication of a change proves the change is live. It does not prove it was crawled, indexed, or read.

Corrections

If a figure, date or claim on this page is wrong, it gets corrected here and logged on the corrections page with what it said, what is true, and why it happened. Entries are never edited away to make the record look better. Corrections are made whether or not anybody asks.

Nothing has been corrected on this page yet. The full log of corrections made anywhere on this site is on the methodology and corrections page. Found something wrong here? hello@robotvisible.com.