How AI Search Picks Sources: What Providers Actually Document

Every week brings another claim about "GEO" (generative engine optimization): add this schema, cut pages into chunks, publish an llms.txt, get mentioned everywhere. Some of it matches what search providers publish. Much of it is a guess presented as a ranking factor.

This guide maps what Google, OpenAI, Perplexity, Anthropic, Microsoft and Apple document about how they crawl your pages, how they pick and cite sources (very little), and which controls you have. You get a provider table, a preview-tag cheat sheet, a claim check, a measurement recipe and a checklist.

Three labels for every claim

Documented means a vendor page states the behaviour or control. Advised means a vendor recommends a practice but publishes no measured effect. Not documented means no vendor has said anything, so whatever you read elsewhere is inference.

We opened every vendor page on 16 October 2026. OpenAI's Help Center refused automated fetches, so its two Help Center pages were read from Internet Archive captures dated 4 and 11 October 2026. Dates below are each page's own "updated" date where it shows one.

Provider by provider: crawling, selection and controls

Three things stand out.

Google now has an explicit switch. Search Console's Search generative AI control, under Settings, lets a property exclude itself from AI Overviews, AI Mode and Discover's generative features. Google says it reached all websites on 31 August 2026 (help page). Advice written before then points to Search's ordinary controls as the only levers; the AI crawlers guide covers the robots.txt side.

Selection is the gap. Google points to its ordinary ranking systems, Microsoft gives unmeasured advice and Apple lists search factors; none gives weights. OpenAI, Perplexity and Anthropic document mostly crawler access.

Citations are not guaranteed. OpenAI's help page says responses "may include" citations, which "can be incomplete, outdated, or incorrect".

Preview controls: what each tag does in each system

Preview tags limit how much of a page a provider may quote. Google and Microsoft document them for AI answers, and Apple documents nosnippet. OpenAI, Perplexity and Anthropic document none in the pages we read.

<!-- Whole page: no snippet, no direct input to Google's AI answers -->
<meta name="robots" content="nosnippet">

<!-- Whole page: at most 160 characters of text (Google, Bing) -->
<meta name="robots" content="max-snippet:160">

<!-- One block only (Google, Bing) -->
<div data-nosnippet>Subscriber-only pricing table</div>

<!-- Bing only: keep the page out of Copilot links -->
<meta name="bingbot" content="noarchive">

<!-- Apple only: opt out of AI-generated answers -->
<meta name="applebot" content="nosnippet">

Use one of the first two; Google applies the stricter rule when both appear. Prefer data-nosnippet for paid or sensitive blocks, because a page-wide nosnippet also changes your ordinary results.

Warning: A crawler can only obey a tag it can fetch. Google says its tags work only on crawlable pages, and Bing and OpenAI say the same for noindex, so do not also block the page in robots.txt.

Which GEO claims have vendor backing

We found no vendor-confirmed ranking factors for AI answers. Google's guide warns that no third-party tool can see its internal ranking or AI systems, which applies to any "AI visibility score".

The academic study behind the term. GEO comes from Aggarwal et al. (KDD 2024), by authors at IIT Delhi, Princeton and independent researchers, with National Science Foundation support acknowledged. They built a test engine (five Google results fed to gpt-3.5-turbo), a 10,000-query benchmark and their own visibility metrics, partly scored by a language model. Adding citations, quotations or statistics raised their main visibility metric by 30 to 40 percent (relative) in that testbed. A Perplexity check used 200 queries with source text uploaded as files, not the live web. We found no commercial conflict declared, but this shows what a 2023-era simulation rewarded, not how any provider ranks today.

Vendor-run counts. xFunnel, a company that tracks performance across AI engines, posted in February 2025 an analysis of 40,000 answers and 250,000 links from ChatGPT, Perplexity and Gemini, using questions it wrote itself. It reports citations by Moz Domain Authority band, but with no baseline of how many web domains sit in each band, that cannot show a preference (our reading).

Measure AI visits, impressions and citations

Four different things get called "AI traffic", and none replaces another.

1. Crawler requests (access logs). Use the server-log guide. A hit from OAI-SearchBot means a page was fetched for an index, not cited.

2. Clicks from AI tools (analytics). In GA4, open Reports, then Acquisition, then Traffic acquisition. The default "AI Assistant" channel covers arrivals "from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok"; AI Overviews and AI Mode count as Organic Search instead (definitions). It names neither Perplexity nor Claude, so also switch to Session source / medium and look for them. OpenAI says ChatGPT adds utm_source=chatgpt.com (Help Center), which also appears in your access log. On a four-line sample log we invented, these printed the landing pages, then the referrer hosts:

   grep 'utm_source=chatgpt.com' access.log | awk -F'"' '{split($2,r," "); sub(/\?.*/,"",r[2]); print r[2]}' | sort | uniq -c | sort -rn
   awk -F'"' '{print $4}' access.log | grep -oE 'https?://[^/]+' | sort | uniq -c | sort -rn | head
   

3. Impressions in Google (Search Console). The Generative AI performance report counts impressions in AI Overviews and AI Mode by page, country, date and device. It may be missing if a site has too few impressions or has excluded itself. The same appearances also sit in the main Performance report, mixed with ordinary results.

4. Citations in Bing and Copilot (Bing Webmaster Tools). AI Performance shows total citations, average cited pages, a sample of grounding queries, per-page citations and a trend, across Copilot, Bing summaries and "select partner integrations" (announcement, Feb 2026). Microsoft says counts do not indicate ranking or placement; clicks are not listed.

5. Read every number as a floor. Microsoft Clarity's AI channel counts only the standalone sites of Claude, Copilot, Gemini, ChatGPT and Perplexity (docs), and a visit with no referrer cannot be attributed. Write one monthly line: crawler hits, AI-assistant sessions, Google impressions, Bing citations.

AgentReady logs hits from recognised AI crawlers by bot and by day, and hosts an llms.txt with Markdown mirrors. That covers step 1 only, and the method works without it.

Owner's checklist: what you can and cannot control

• Confirm pages you want cited are indexable, return 200, and carry no noindex or nosnippet.
• Read the live /robots.txt for OAI-SearchBot, PerplexityBot, Claude-SearchBot, Bingbot and Applebot. The AI crawler robots.txt generator builds an RFC 9309 file in your browser, with each bot linked to its vendor page.
• Allow the OpenAI and Perplexity published IP ranges through your CDN or firewall; both vendors ask for it.
• Decide training separately: GPTBot, ClaudeBot, Google-Extended, Applebot-Extended.
• Check that Search Console shows the generative AI choice you intend; the default is include.
• Use data-nosnippet for blocks you will not have quoted.
• Submit a sitemap with accurate lastmod to Bing and use IndexNow.
• Give each URL one topic, key facts near the top, and no duplicates.
• Set up the GA4 view and log recipe, then record the monthly line.
• Re-read vendor pages each quarter.

Outside your control: which pages a provider retrieves, how many it links and in what order, whether an answer cites anyone, which partner provider ChatGPT queries, and whether a click carries a referrer.

The short version

Every provider documents crawler access; none publishes a selection formula. Spend effort where the documents agree: crawlable, indexable, clearly written pages; deliberate preview tags; the Search Console setting; and a monthly measurement line. Treat any other tactic as a test, not a rule.

Frequently asked questions

Is GEO different from SEO?

Google says optimizing for its generative AI features is still SEO, because they rely on Search's ranking systems. Microsoft defines GEO as work on eligibility for grounding and citations, and says it guarantees neither. OpenAI, Perplexity and Anthropic publish no ranking recipe, so no vendor documents a separate discipline.

Can I stay in Google Search but opt out of AI Overviews and AI Mode?

Yes. Open Search Console, then Settings, then Search generative AI, and choose Exclude. Google says it generally takes a few days, you get no impressions or traffic from those features, and AI training is unaffected, which needs Google-Extended. To remove the page from Search entirely, use noindex.

How do I see visits from ChatGPT, Perplexity or Copilot?

In GA4 Traffic acquisition, check the AI Assistant channel and search Session source for chatgpt.com, which OpenAI says it adds as utm_source. We found no Perplexity documentation of referral marking, so list referrer hosts in your access log. Clicks from Google's AI Overviews and AI Mode appear under Organic Search.

Will llms.txt, schema or chunking get me cited?

No provider documents that they do. Google says Search ignores llms.txt, needs no special markup and has no chunking requirement. Microsoft says structured data may support clearer grounding, with no guarantee. Do them for other reasons.

Sources and further reading

• AI features and your website - Google Search Central, updated 10 December 2025. Eligibility, controls, measurement.
• Optimizing your website for generative AI features on Google Search - Google Search Central, updated 10 July 2026. Retrieval, "GEO" myths, third-party tools.
• Robots meta tag, data-nosnippet, and X-Robots-Tag specifications - Google Search Central, updated 24 March 2026. Snippet tags and AI answers.
• Search generative AI control - Search Console Help, accessed October 2026. Include, exclude, timing.
• Publishers and Developers - FAQ - OpenAI Help Center, Internet Archive capture of 4 October 2026. OAI-SearchBot, noindex, referral parameter.
• Bing Webmaster Guidelines - Microsoft Bing, accessed October 2026. Grounding advice, noarchive, nocache.
• About Applebot - Apple Support, published 4 September 2026. AI-answer context, nosnippet, ranking factors.
• GEO: Generative Engine Optimization - Aggarwal et al., KDD 2024. Design and limits of the GEO study.