Contents
In short. One question typed into a chatbot proves nothing, because AI answers change from run to run. Measure a share instead. Keep a fixed set of prompts, run it many times in fresh sessions, and count in how many answers your company is named and in how many your website is linked. Then compare the result with the official reports from Google and Bing and with referral visits in your analytics. None of these gives the full picture, so read them side by side rather than adding them into one score.
AI visibility is three separate facts, not one score
AI visibility is three observations: crawlers open your pages, an answer names or links you, people click and enquire. Each has its own tool; one "visibility score" hides more than it shows. Our article on preparing a site for search and AI answers covers what to build; this one, how to check.
Disclosure: the publisher, WEBNOSTRA CAPITAL, is connected to WEBNOSTRA, which sells SEO and GEO services (optimising for generative answer systems). You can run this test yourself with no agency or paid software; a service link appears only at the end.
Scroll right to see all columns
| Layer | Where to look | What it does not prove |
|---|---|---|
| Access and indexing | Search Console URL Inspection, Bing Webmaster Tools, server logs | That the page will be shown |
| Visibility in answers | Platform reports and your own test | That anyone visited |
| Business result | Analytics, UTM tags, CRM | That the AI answer was the cause |
Indexed does not mean shown in an answer
If the page is indexed but absent from answers, the cause may be content, the prompts, the system itself or chance.
Indexing means a search engine stored the page; visibility means an answer used it and showed a name or link. Google’s guide to AI features in Search (updated 10 July 2026) requires a page to be indexed and snippet-eligible, yet "Indexing and serving aren’t guaranteed". AI Overviews and AI Mode have no extra technical requirements (AI features page, 10 December 2025).
URL Inspection in Search Console shows the indexed version and tests indexability, but not whether an answer named you. If crawlers cannot reach the page, fix access first. If the page is indexed but absent from answers, the cause may be content, the prompts, the system itself or chance.
Which crawlers to let in
A crawler is a program a platform uses to visit sites; you set rules for it in robots.txt. For AI search, allow the search crawlers of Google, Bing, OpenAI, Perplexity and Anthropic; decide on training crawlers separately, since per OpenAI and Anthropic they are distinct from search. Take exact strings and addresses from the platforms: they change.
An allowed crawler is a precondition, not a promise of citation. An editorial proposal for a one-evening check:
- Open robots.txt and confirm none of the search crawlers in the table is disallowed.
- Review bot protection in your CDN and web application firewall (WAF): they can block crawlers regardless of robots.txt.
- Find crawlers in server logs and check addresses against searchbot.json (OpenAI), perplexitybot.json, bots.json (Anthropic) and bingbot.json. We assume an address is harder to fake than a name.
- After a robots.txt change, wait: OpenAI and Perplexity warn of a delay of up to about 24 hours.
Scroll right to see all columns
| Crawler | Platform and purpose | What blocking does, per documentation |
|---|---|---|
| Googlebot | Google Search, all features | No index, so not eligible for Search AI features |
| Bingbot | Bing, Copilot, Bing AI summaries | Bing’s crawler page was unreadable to us; the effect is our assumption |
| OAI-SearchBot | ChatGPT search; OpenAI recommends allowing it | No ChatGPT search answers, except navigational links |
| GPTBot | OpenAI training | Content not to be used for training; independent of search |
| ChatGPT-User | OpenAI, user actions | robots.txt may not apply; no effect on search |
| PerplexityBot | Perplexity search, not training | Perplexity recommends allowing it |
| Perplexity-User | Perplexity, user actions | Generally ignores robots.txt |
| Claude-SearchBot | Claude search | No indexing for search; may reduce visibility |
| Claude-User | Pages a Claude user needs | May reduce visibility |
| ClaudeBot | Anthropic training | Future material excluded from training |
What the official reports show, and what they cannot
Google’s and Bing’s reports count different events: never add or directly compare them. Each sees one platform; for ChatGPT, Perplexity and Claude we found no such report for site owners.
Google announced Search Generative AI performance reports on 3 June 2026; a note says all sites had them by 31 August, and a missing report usually means too few impressions. Bing’s AI Performance, in public preview since 10 February 2026, covers Copilot, Bing AI summaries and partner integrations; grounding queries are the phrases the AI searched with.
For readers in Russia: Yandex Webmaster reports Share of Voice in Alice AI answers (three months, weekly), only where the site already ranks high, with no clicks. The answers come from pages Yandex has indexed; add Alice AI if your audience uses Yandex.
Scroll right to see all columns
| Report | What it shows | What it omits or limits |
|---|---|---|
| Google Search Console: Generative AI performance | Link impressions in AI Overviews and AI Mode by page, country, device, date | No query dimension, no Search Labs data; clicks sit in the Performance report |
| Bing Webmaster Tools: AI Performance | Citations in total and by page, cited pages per day, grounding queries, trend | No ranking, importance or placement; grounding queries are a sample |
How to run the test: a fixed set, many runs, two shares
The protocol is an editorial proposal, not a standard. It rests on the SparkToro and Gumshoe finding that answers are unstable (last section) and on common sense: more repeats, less chance.
Workload: 20 prompts, 10 runs and 3 platforms give 600 answers, 20 hours at an assumed two minutes each. Start with one platform or 5 runs; 10 is a compromise, not a proven threshold. A new site should expect zeros and watch indexing and official reports.
- Write 20 prompts in four families of 5: definitions, comparisons, "how to calculate", choosing a supplier; some name your company, some do not. Freeze the set.
- Pick two or three platforms your buyers use: Google (AI Overviews and AI Mode), ChatGPT with search, Perplexity, Claude with search, Bing Copilot.
- Run each prompt 10 times within 1-3 days in fresh sessions: new conversation, same search mode, country and language, memory off.
- Log a row per run: date, platform, prompt ID, run, company named in the text (1/0), your domain linked (1/0), competitors. Flag errors; never delete them.
- Count two shares, mentions and links; SparkToro calls an "average position in AI" uninformative. Repeat every two to four weeks.
A rise from 20% to 26% is not yet growth
Count something as a change only if it repeats in two consecutive rounds and shows in several prompt families.
Illustrative example, invented numbers: one platform, 20 prompts, 10 runs, 200 answers. Round one: the name in 40 answers (20.0%), a link in 10 (5.0%). Four weeks later: 52 (26.0%) and 16 (8.0%).
The 95% Wilson intervals (table below) overlap, so growth is not shown. For one prompt with 3 mentions in 10 runs (30%) the interval is 10.8-60.3%: treat single prompts as examples and decide from families and the whole set.
Answers to one prompt are related, so real uncertainty is larger. Editorial proposal: count something as a change only if it repeats in two consecutive rounds and shows in several prompt families.
| Item | Round 1 | Round 2 |
|---|---|---|
| Company named | 20% | 26% |
| Link to your site | 5% | 8% |
Wilson intervals overlap (table below): growth not shown. Numbers invented for the calculation.
Scroll right to see all columns
| Measure | Round 1 | Round 2 | Reading |
|---|---|---|---|
| Company named in the answer | 40, 20.0% (15.0-26.1) | 52, 26.0% (20.4-32.5) | Growth not shown |
| Link to your domain | 10, 5.0% (2.7-9.0) | 16, 8.0% (5.0-12.6) | Growth not shown |
Visits from AI are only partly visible in analytics
Treat the figure as a floor and reconcile it with the CRM. A referrer tells where a visitor came from; UTM tags are link labels read as the source.
Google Analytics help describes an AI Assistant channel: visits from ChatGPT, Gemini, Deepseek, Copilot, Grok and similar, medium "ai-assistant" when the referrer matches a list. It excludes AI Overviews and AI Mode: judge those in Search Console.
Per a Search Engine Journal article of 15 July 2026 (a trade publication, not verified in Google’s help), the channel was added on 13 May 2026, reached most properties by June, its list changes, and Perplexity is not on it. Editorial proposal:
- Build a GA4 custom channel group by source (chatgpt.com, perplexity.ai, claude.ai, copilot.microsoft.com, gemini.google.com) above Referral; per SEJ it applies to past data too.
- Treat the result as a lower bound.
- Add "How did you hear about us?" to your form and reconcile the answers with the CRM.
- Judge return on investment over a period with enough enquiries, not from a few dozen visits.
Scroll right to see all columns
| Situation | Where the visit lands | Basis |
|---|---|---|
| Link with no tag and no referrer | Direct | SEJ article; Google’s help: Direct means a saved link or typed address |
| In-app browser stripped the referrer | Source may survive, medium is lost | SEJ article |
| Platform not on the AI Assistant list | Referral, as with Perplexity | SEJ article; the list changes |
| Click from AI Overviews or AI Mode | Google organic search | GA4 help: the channel excludes them |
| Saw the name, searched later | Organic search or Direct | Editorial inference |
What this test cannot prove
The test shows a share within your prompt set on the days you ran it, not a market share and not the cause of any sales growth.
The SparkToro and Gumshoe study (Search Engine Journal, 30 January 2026): 2,961 prompts to ChatGPT, Claude and Google AI, 12 brand-recommendation prompts run 60 to 100 times per platform by volunteers. The same brand list appeared in fewer than 1% of runs, in the same order in fewer than 0.1%. The authors trust a visibility percentage over a position.
No source we checked says how often people ask AI your prompts. Appearance is never guaranteed, Google says, and no third-party tool can access its ranking or AI systems. The protocol numbers are an editorial proposal, and linking growth to your actions takes separate observation of enquiries.
Key decisions
- separate three facts: the page is reachable and indexed, the company is named or cited, and people click through and enquire
- measure the share of answers that mention you or link to you across many repeated runs, not a position in one answer
- read official reports and analytics as different observations of different events, and state the limits of the method
What to discuss with your team
- A set of 20-30 prompts in four families is frozen
- Runs and fresh-session conditions are written in the log
- Two shares are counted, mention and link, with no AI position
- Robots.txt, CDN and WAF are checked and crawler addresses compared with published lists
- Google and Bing reports are read separately and never added together
- Visits from AI are a floor, reconciled with CRM enquiries
Terms in plain language
- Run
- One execution of a prompt in a new platform session; shares are computed over many runs because a single answer is partly random
- Mention and citation
- A mention is the company name in the answer text; a citation is a link to your domain among the answer’s sources. They are different events and are counted separately
- Wilson interval
- A way to estimate the range a share could lie in when you have a limited number of answers; fewer answers means a wider range
Sources
- Google Search Central: Optimizing your website for generative AI features on Google Searchdevelopers.google.com · checked 31/10/2026
- Google Search Central: AI features and your websitedevelopers.google.com · checked 31/10/2026
- Search Console Help: Generative AI performance report (Search)support.google.com · checked 31/10/2026
- Google Search Central Blog: Introducing Search Generative AI performance reports in Search Console, 3 June 2026developers.google.com · checked 31/10/2026
- Google Search Central: Google common crawlers, including Googlebotdevelopers.google.com · checked 31/10/2026
- Search Console Help: URL Inspection toolsupport.google.com · checked 31/10/2026
- OpenAI: Overview of OpenAI crawlers (OAI-SearchBot, GPTBot, ChatGPT-User)developers.openai.com · checked 31/10/2026
- Perplexity: PerplexityBot and Perplexity-User crawlersdocs.perplexity.ai · checked 31/10/2026
- Anthropic Support: ClaudeBot, Claude-User and Claude-SearchBot crawlerssupport.claude.com · checked 31/10/2026
- Bing Webmaster Blog: Introducing AI Performance in Bing Webmaster Tools Public Preview, 10 February 2026blogs.bing.com · checked 31/10/2026
- Yandex Webmaster: Site visibility in Alice AI (Russian-language page)yandex.ru · checked 31/10/2026
- Yandex Webmaster: Search with Alice, YandexAdditional and YandexAdditionalBotyandex.ru · checked 31/10/2026
- Google Analytics Help: Default channel group, AI Assistant channelsupport.google.com · checked 31/10/2026
- Search Engine Journal: AI recommendations change with nearly every query, SparkToro (30 January 2026)searchenginejournal.com · checked 31/10/2026
- Search Engine Journal: How to track AI traffic in GA4 without undercounting it (15 July 2026, trade publication)searchenginejournal.com · checked 31/10/2026