The Ten Commandments of AI Visibility

Lighthouse also checks whether forms are exposed to the browser's built-in agent through WebMCP, which is still in an origin trial.

Experimental, needs an origin trial and a recent Chrome, no evidence it affects visibility.

Bing Webmaster Tools now reports how often your pages are cited as sources in Microsoft Copilot and Bing's AI summaries: total citations, cited pages per day, a sample of the questions, citations per page, and a trend.

Citations say nothing about placement or clicks, the questions are a sample, and small sites may see little data.

Checking citations through OpenAI's API and Google results together recovers only about a quarter of the pages the ChatGPT app actually cites, and the cited pages vary a lot from one run to the next.

Preprint, not peer reviewed. 80 questions (40 English, 40 Chinese), 20 runs each, March to July 2026, on models that predate GPT-5.6 and GPT-6. 40.8 to 44.5% of distinct pages appeared in only one of the 20 runs.

The same question asked in different languages can lead an AI assistant to recommend a different business, and switching web search on changes the winner in most cases.

A practitioner test by SALT.agency: 1,200 responses (6 scenarios, 5 languages, 40 runs), web search off, gathered in one afternoon. Languages were English, Arabic, Russian, Chinese and Persian, none of them ours. The author calls the results suggestive.

ChatGPT has shifted its citations toward official and primary sources such as government, research, regulator and brand sites, and away from explainer and aggregator sites. It also cites Reddit and Wikipedia far less than it did.

Ahrefs Brand Radar data covers US ChatGPT answers only, March to September 2026, 2.4 to 3.8 million answers a month. The official-source test used 20,000 US prompts: 84.4% of factual answers cited at least one official site. Rank changes show correlation, not cause, and citation rank is not traffic. Reddit went from about 19.1% of answers to 2.3%, and Wikipedia from 14.1% to 2.9%.

When ChatGPT searches the web, it rewrites the question into new searches that more and more often add the current year, the word 'official', or 'price'. But only about 1 in 7 answers triggers a search at all.

Sample: 66.6 million free, logged-out GPT-5.6 Luna responses and 15.0 million fan-outs, August to September 2026. 14.4% of responses searched, averaging about 1.6 fan-outs per search. Shares in September: '2026' 22.9%, 'official' 13.1%, 'price' 8.9%. Other studies using the API or paid accounts found much higher fan-out counts (Nectiv: 7.61 per prompt). OpenAI changes retrieval often.

About 7% of ChatGPT's searches are restricted to one website ('site:'). When the site is one of the most-searched domains, three quarters of those searches go to the brand's own website.

6.8% of 15 million fan-outs, spread across about 116,000 domains. The 75.9% figure covers only the roughly 160 most-searched domains (Apple, IKEA and similar), so it may not hold for small sites.

Google AI Overviews now appear on most searches for a brand name, and when sources are listed they lean heavily on Wikipedia, YouTube and LinkedIn.

The 82.91% figure is one day (29 September) of US desktop crawls: 92.5 million branded crawls out of 232.8 million. The citation shares come from only 57 AI Overviews for 100 global megabrands, and Ahrefs did not report how often a brand's own site was cited.

The searches Gemini runs behind the scenes are machine-written, long (about 6.5 words), often stuffed with terms or operators, and a third mention a year. They do not match what people type into Google.

Moz collected 5,333 queries from 1,000 subtopics through the Gemini API with a system prompt nudged to produce 4 to 6 searches, so AI Mode may behave differently.

A Moz piece argues that AI agents use sites the way screen readers do, and recommends real buttons and links with specific labels and facts that agree on every page.

The author works for a vendor and offers no data.