Whether assistants cite you cannot be read in any report: not Search Console, not the vendors, not our test. The only honest measurement is a protocol you run yourself: a fixed set of questions, asked to the same assistants at a regular interval, recorded the same way every time. This page describes the method without publishing our question list.
1. Why the test does not measure citations
What the vendor documents. Google states that "there are no additional requirements to appear in AI Overviews or AI Mode", that traffic from these features is "included in the overall search traffic in Search Console", within the "Web" type of the Performance report, without a separate breakdown, and that nosnippet, max-snippet and noindex limit what is shown (Google, AI features in Search). No assistant vendor publishes a per-site citation report.
What we recommend. Do not confuse prerequisites with results. The GEO test scores what makes a citation possible, AI answer crawlers allowed, text readable without JavaScript, identity entity, and its methodology says it: it does not measure your actual citations in assistant answers, nobody can reliably today. A tool that displays a "citation rate" without saying which questions it asked, to whom, when and how many times, displays a number without a source.
Check in five minutes whether your business appears
Before any protocol, a quick check: open ChatGPT in search mode, in a fresh session, and ask three questions a client would ask without your name, by need, by city, by comparison. Note whether you are cited, which page, and who is cited instead; do the same on Perplexity. This is not a measurement, it is a snapshot: it says whether the subject is absence or position, and it is the starting point of the protocol below.
2. A fixed prompt set
What the vendor documents. OpenAI documents that OAI-SearchBot is "used to surface websites in search results in ChatGPT's search features" and that sites that opt out "will not be shown in ChatGPT search answers": what is observable is a site appearing in a search answer, answer by answer (OpenAI, bots and crawlers).
What we recommend. A question set written once, frozen, and asked as is. The questions are the ones a client would ask before looking for you: by need, by city, by comparison, by problem, never with your name. Twenty to forty questions are enough for a services site; beyond that the run gets too long to repeat. Each question is asked to every assistant retained, in a fresh session, without history or personalisation, in search mode when the assistant offers one. We do not publish our list: a published list becomes an optimisation target and stops measuring.
How to fix
The protocol fits in a definition file, versioned, that only changes by adding a version.
# protocol v1 - frozen on 2026-09-08
assistants: [chatgpt-search, perplexity, claude, google-ai-mode]
frequency: monthly, first week of the month, same weekday
session: new, no history, no personalised account, language en
questions: 28, split: need (10), city (6), comparison (6), problem (6)
# the questions themselves live in a private file: questions-v1.txt3. What to record
What the vendor documents. Anthropic documents that Claude-User "supports Claude AI users. When individuals ask questions to Claude, it may access websites"; a question asked during the run can therefore trigger a request visible in your logs, under that name, from the published addresses (Anthropic, web crawling).
What we recommend. One line per question and per assistant, with the same columns on every run: the date, the protocol version, the assistant, the question id, whether your site is cited, the exact URL cited, its position among the sources, the competitors cited, and a screenshot of the answer. Then, within the hour, a look at the server logs to note whether a fetcher came for the cited page. The screenshot is what makes the measurement verifiable by someone other than you.
How to fix
date;protocol;assistant;question_id;cited;cited_url;position;competitors_cited;screenshot
2026-09-08;v1;chatgpt-search;Q07;yes;https://your-site.com/en/guides/x;2;competitor-a.com,competitor-b.com;2026-09-08-chatgpt-Q07.png
2026-09-08;v1;perplexity;Q07;no;;;competitor-a.com;2026-09-08-perplexity-Q07.png4. Which indicators to follow, and how to read them
What the vendor documents. Google warns that its preview controls "may take time to process", because "crawling can take anywhere from several days to several months" (Google, AI features in Search). A change on your site is not visible in an answer the next day.
What we recommend. Compare runs, never answers. An answer varies from one session to the next; a run of thirty questions across four assistants smooths that variation. Three useful readings: the share of questions where you are cited, per assistant; the questions where a competitor is cited and you are not, which name the pages to write; and the page cited, which says what the assistant treats as your best answer. A fix is judged over at least two runs after it goes live, and a swing of one or two questions is not a signal. We have no benchmark figure to give you: we only publish what we measured, and what we measured belongs to our clients.
What we did on seoforge.fr
We have no citation measurement tool, and our test does not claim to be one: its page says so among its limits. The measurement described here is the one we run by hand during a GEO Audit, before and after the fixes, and it is delivered to the client with its screenshots.
Going further
This page details the "what GEO does not guarantee" section of the AI search optimisation guide. If your business appears in no answer at all, the step-by-step diagnosis walks up the chain of causes. The GEO Audit includes the measurement described here, before and after.
Frequently asked
Which tools measure AI visibility? Tracking tools exist and round-ups compare them; we name none we have not used. The protocol on this page needs none: a question set, four assistants, a spreadsheet and screenshots are enough for an SME.
Which KPIs for GEO? Three, read across runs rather than on single answers: the share of questions where you are cited, per assistant; the questions where a competitor is cited and you are not; the page cited. The technical prerequisites are read in the GEO test.