Skip to content

AI Search Ranking Factors: 10 Signals Ranked by Evidence

The 10 factors that decide whether AI engines cite your content, ranked by evidence strength, with the study behind each one and where to start.

Your CEO forwards a screenshot of ChatGPT recommending three competitors in your category and asks whether ChatGPT or Perplexity ever recommends the company. Most content teams can't answer. Their dashboards track rankings and clicks, and an AI answer produces neither unless someone clicks a citation.

AI engines cite pages they're allowed to crawl, passages that answer one question on their own, and brands the rest of the web already discusses. The 10 signals below are ranked by the strength of the evidence behind each one, not by how hard vendors market them, and the final section turns them into a sequenced checklist for how to optimize for AI search.

How AI search engines choose which sources to cite

AI search engines first retrieve candidate pages through search, then write an answer and attach citations to the passages that shaped it. A page can be retrieved and still never cited, so each stage needs its own work.

AI visibility is the measure of that outcome. AI visibility is how often answer engines name your brand and cite your pages for the questions your buyers ask, and how they describe you when they do. A mention is your brand named in the answer text. A citation is your URL listed as a source. The two travel separately, and a page can supply a citation without the brand ever being named. Our guide to how AI engines choose citations covers what happens between retrieval and the final answer in more depth.

Retrieval-augmented generation in plain terms

Retrieval-augmented generation (RAG) means the model looks things up before it answers. The engine runs searches, pulls back pages, splits them into passages, scores those passages against the question, and writes a response grounded in the best matches.

Engines rarely search your buyer's exact words. Google documents that AI Overviews and AI Mode may use query fan-out, issuing several related searches across subtopics and data sources while the response is built. Google's own example splits a question about fixing a weed-filled lawn into separate searches about herbicides, chemical-free removal and prevention. OpenAI describes ChatGPT search the same way, rewriting a prompt into one or more targeted queries that it sends to search providers.

Retrieval systems score chunks of a page, not whole pages. A chunk that depends on a paragraph three screens up loses the context that made it relevant. A passage that names its subject and answers one sub-question in its first sentence can be scored on its own for any fan-out query it matches. That last point is our inference from how chunked retrieval works rather than something Google or OpenAI document, but it's the most useful mental model we know for writing pages engines quote.

How ChatGPT, Perplexity, Gemini and Google AI Overviews differ

Each engine retrieves from a different place and displays sources differently. This comparison draws on each company's current documentation:

EngineWhere it retrievesHow it shows citationsWhat to optimize
ChatGPT searchPages surfaced by OAI-SearchBot plus third-party search providers, with prompts rewritten into targeted queries. OpenAI doesn't name its consumer search partners; it names Bing for Enterprise and Edu queriesInline citations plus a Sources list at the end of the responseAllow OAI-SearchBot and OpenAI's published IP ranges, and answer near the top of the page
PerplexityPerplexity's own web index built by PerplexityBot, searched in real time for each questionNumbered inline citations linking to the original sources on every answerAllow PerplexityBot, keep pages current, and make each passage self-contained
GeminiGoogle documents Grounding with Google Search for the Gemini API; its official documentation doesn't describe the Gemini app's retrieval pathA Sources button or inline links, though not every response includes sourcesGoogle indexing, plus a decision on Google-Extended, which governs grounding in Gemini Apps
Google AI Overviews and AI ModeGoogle's Search index, using query fan-out across subtopicsSupporting links inline and beside the answer, with more inline links added in October 2026Indexed, snippet-eligible pages that answer the sub-questions fan-out generates

Perplexity puts numbered citations on every answer, which makes it the easiest engine to audit by hand. Our Perplexity citation playbook covers it in detail.

How AI search optimization differs from traditional SEO

Most of what works in SEO carries over. Google's documentation for its AI features lists the same foundational practices it recommends for regular search, and a page has to be indexed and eligible for a snippet before it can appear as a supporting link in AI Overviews or AI Mode.

The vocabulary is newer than the practice. Answer engine optimization (AEO) is the work of making pages and third-party signals citable by AI answer engines. Generative engine optimization (GEO) is the name a 2024 research paper gave the same goal, and large language model optimization (LLMO) is a third label for it. We use AEO for the work and AI visibility for the outcome you measure.

These four things change when the goal is citation rather than rank:

  • The unit of competition shrinks from the page to the passage, because engines score chunks.
  • Unlinked brand mentions on other websites carry weight that classic link building ignored.
  • Success shows up as mentions and citations inside answers, which rank trackers can't see.
  • Each engine retrieves differently, so a page cited by Perplexity may never appear in ChatGPT for the same question.

Crawlability, indexing, rankings and useful content still do most of the lifting. We've written separately about where AEO genuinely departs from SEO if you need to make that case internally.

The 10 signals at a glance

The table below sorts each signal by how strong its supporting evidence is. Strong means documentation or large-sample data directly supports it, moderate means the signal correlates with citation or follows from engine mechanics without an isolating study, and emerging means early or mixed evidence:

SignalEvidence strengthKey source
1. Crawler access for AI botsStrongOpenAI, Anthropic, Perplexity and Google crawler documentation
2. Answer-first, chunked passagesStrongGrowth Memo analysis of 18,012 ChatGPT citations, February 2026
3. Third-party mentions and roundup inclusionStrongMuck Rack, 207 million AI citations collected in 2026
4. Existing search rankings and index presenceStrongGoogle eligibility rules and Ahrefs data on 4 million AI Overview URLs, January 2026
5. Topical authority through Content ClustersModerateGoogle query fan-out documentation
6. E-E-A-T and author credibilityModerateGoogle helpful-content guidance
7. Content freshnessModerateSeer Interactive, 47,097 citations, March to June 2026
8. Schema markup and structured dataModerateAhrefs matched test of 1,885 pages, August 2025 to March 2026
9. Original data, statistics and cited sourcesEmergingGEO paper, KDD 2024, 10,000 queries
10. llms.txt and AI-specific filesEmergingAhrefs study of 137,000 domains, May 2026

Signals with strong evidence

These four come first because each one either gates eligibility outright or rests on large datasets from multiple engines. Get them right before spending time on anything lower in the table.

1. Crawler access for AI bots

Crawler access is the one signal on this list that works as a hard gate. If an engine's search crawler can't fetch a page, nothing else here matters for that engine.

The major AI companies now separate crawlers that index for search from crawlers that collect training data, and each robots.txt rule works independently.

  • Websites that block OAI-SearchBot won't be shown in ChatGPT search answers, though they can still appear as navigational links, while blocking GPTBot only signals that content shouldn't be used to train OpenAI's foundation models.
  • Anthropic splits the same way. Blocking ClaudeBot excludes future content from training, and blocking Claude-SearchBot or Claude-User reduces visibility in Claude's search results.
  • PerplexityBot powers Perplexity's search results and isn't used for training. When it's blocked, Perplexity won't index the text, though it may still keep the domain, headline and a short factual summary.

Google works differently. Google-Extended has no crawler of its own. Google-Extended is a robots.txt token that controls whether content Googlebot already crawls trains Gemini models and grounds answers in Gemini Apps, and Google states it doesn't affect inclusion or ranking in Google Search.

Appearance in AI Overviews and AI Mode depends on indexing, the nosnippet and max-snippet rules, and a separate generative AI control in Search Console. Google's October 30, 2026 announcement adds that websites allowing Google-Extended are eligible to appear in Gemini Apps answers through grounding, so blocking it carries a Gemini visibility cost.

For B2B teams that want AI visibility but prefer to keep content out of training sets, this pattern is a sensible default:

# Search and answer crawlers: allow
User-agent: OAI-SearchBot
Allow: /

User-agent: Claude-SearchBot
Allow: /

User-agent: PerplexityBot
Allow: /

# Training crawlers: block
User-agent: GPTBot
Disallow: /

User-agent: ClaudeBot
Disallow: /

# Optional: blocking also removes Gemini Apps grounding
# User-agent: Google-Extended
# Disallow: /

Robots.txt is only half the gate. OpenAI says the host or CDN must also allow traffic from its published search-bot IP ranges, and Perplexity warns that Bot Fight Mode can still challenge its crawlers. The user-initiated fetchers, ChatGPT-User and Perplexity-User, act on a person's request and may not follow robots.txt at all. Check the firewall rules, then allow about 24 hours for OpenAI and Perplexity to reflect a robots.txt change.

2. Answer-first, chunked passages

In an analysis of 1.2 million ChatGPT responses and 18,012 verified citations, published by Growth Memo in February 2026, 44% of citations came from the first 30% of a page's text. A July 2026 Search Engine Land analysis of 15.7 million AI Mode citations across 148 industries found about 85% of highlighted passages were fully self-contained.

Both findings point to the same writing habits:

  • Put the direct answer in the first sentence under each heading, then support it.
  • Name the subject in every section instead of leaning on "it" or "this tool".
  • Keep one idea per chunk, and split a section that answers two sub-questions.
  • Write headings as the sub-questions a fan-out query would generate.
  • Use tables for comparisons a reader would scan, with full facts in each row.

Here's how a typical section changes when you rewrite it this way. Before, the answer arrives late and never names its subject:

Analytics has changed a lot in recent years, and many marketers feel unsure how to handle new traffic sources. There are several approaches teams take. In this section, we'll look at one of them and why it might help.
GrowthX

After, the first sentence answers the question and the second gives the operational detail:

A GA4 custom channel group sorts sessions into channels you define, such as AI assistants, using rules on source and medium. GA4 assigns each session to the first channel whose rule matches, so place the AI channel above Referral.
GrowthX

The rewrite survives being lifted out of the page. An engine retrieving it for a sub-query about GA4 channels gets a complete, citable answer with no setup to discard.

3. Third-party mentions and roundup inclusion

Most of what engines cite lives on other people's websites. More than 84% of links cited in AI responses come from earned and third-party sources. The October 2026 report from Muck Rack, a PR software vendor, draws on more than 40 million AI responses and 207 million citations collected in 2026, plus a September analysis of 1,020 prompts across ChatGPT, Claude, Gemini and Perplexity.

For software brands, that third-party layer includes comparison roundups, review platforms and community discussions. And those mentions don't need a link to count. An unlinked mention in a respected category roundup still puts your name next to the topic in the text an engine reads.

Treat this as a placement program with a target list. Run your category prompts, record which third-party pages each engine cites, and pursue inclusion on those specific pages. Keep review profiles current, because evaluation-stage prompts pull from them. Write for the publications your buyers already read rather than buying placements.

4. Existing search rankings and index presence

Ranking is the entry ticket for Google's AI features and a head start everywhere else. Google's rule is plain. A page must be indexed and eligible to show with a snippet to appear as a supporting link in AI Overviews or AI Mode, with no additional technical requirements.

Ranking raises the odds without guaranteeing a citation. 38% of pages cited in AI Overviews also rank in Google's top 10 for the same query, in an analysis of 863,000 keyword SERPs and 4 million AI Overview URLs from January 2026 Gemini 3 results. In the same analysis, about a quarter of cited URLs ranked 11–100 and over a third ranked outside the top 100. ChatGPT also draws on outside search providers, and OpenAI names Bing as the provider for Enterprise and Edu queries, so Bing indexing deserves a check in Bing Webmaster Tools.

Target the sub-questions around each priority query as well as the main query. A page that ranks for "AI search tracking" and also answers "how to segment AI traffic in GA4" can sit in more of the result sets an engine pulls from.

Signals with moderate evidence

The next four signals correlate with citation in vendor data or follow logically from how engines retrieve, but no controlled study isolates their effect. Do them because they make pages better for readers and engines alike, and don't expect any single one to move citations by itself.

Query fan-out is the mechanism that makes depth pay. When Google splits one question into several searches across subtopics, a set of pages covering those subtopics gives the engine more places to find you. Correlation data adds a related point. Ahrefs' study of 75,000 brands found raw page count has almost no relationship with AI visibility. That result suggests publishing more pages alone is unlikely to help. Build pages that cover the sub-questions your buyers ask.

A topic becomes a Content Cluster when you decide to build authority on it, and that cluster then produces the specific pages worth writing. In practice that's a hub page answering the core question, spoke pages answering each sub-question, and internal links connecting them in both directions with descriptive anchor text.

We'd build Content Clusters even though no published study measures internal linking's effect on AI citations directly. They fix the fan-out coverage problem and help classic rankings, which feed signal 4.

6. E-E-A-T and author credibility

Trust is the part of E-E-A-T Google weighs most. Google's guidance says experience, expertise and authoritativeness all contribute to trust, and that content doesn't have to show every one of them. Google also recommends accurate author bylines wherever readers would reasonably expect them.

On AI-assisted writing, Google is explicit that using AI gives content no special advantage. What counts is whether the page is helpful and reliable, regardless of whether AI or a human wrote it.

No study isolates author credentials as a cause of AI citation. The cheap, defensible version is named authors with real credentials, author pages that link to their published work, and first-hand evidence on the page, such as screenshots, process detail and numbers from your own operations. Those details also give engines specific claims to cite, which generic advice never does.

7. Content freshness

Cited pages skew recent, and the skew varies by engine. Across 7,683 cited pages and 47,097 citations from ChatGPT, Gemini and Perplexity between March and June 2026, the share of dated pages updated within the past year was 78% for Gemini, 73% for ChatGPT and 65% for Perplexity. The same study found pages cited consistently across all four months were older, on median, than pages cited in a single month, so freshness looks more useful for getting into answers than for staying there.

Refresh the substance, not the date stamp. Update the statistics, examples and screenshots on priority pages, then show a visible last-updated date that reflects the change. Refreshing a handful of pages is easy, while refreshing a whole library at real scale needs a queue sorted by traffic direction and business priority.

8. Schema markup and structured data

Schema helps engines understand entities, but the best evidence shows no citation lift from adding it. Google's AI optimization guide says structured data isn't required for generative AI search and that no special markup exists for it. In a matched test of 1,885 pages that added JSON-LD between August 2025 and March 2026, measured against 4,000 control pages, Google AI Mode moved 2.4% and ChatGPT 2.2%, both statistically indistinguishable from zero, while AI Overviews fell 4.6%.

Schema still earns its place. A Microsoft Bing product manager confirmed in March 2025 that Microsoft uses schema to help its language models understand content. Google stopped showing FAQ rich results in Search from May 7, 2026, so treat FAQPage markup as entity description only. Every fact in the markup must match visible text on the page, which is also where engines read it.

This compact graph covers Organization, Article and FAQPage in one block:

{
  "@context": "https://schema.org",
  "@graph": [
    {
      "@type": "Organization",
      "@id": "https://example.com/#org",
      "name": "Example Co",
      "url": "https://example.com",
      "logo": "https://example.com/logo.png",
      "sameAs": ["https://www.linkedin.com/company/example-co"]
    },
    {
      "@type": "Article",
      "headline": "How to segment AI traffic in GA4",
      "author": {"@type": "Person", "name": "Jordan Lee", "url": "https://example.com/authors/jordan-lee"},
      "publisher": {"@id": "https://example.com/#org"},
      "datePublished": "2026-03-02",
      "dateModified": "2026-10-20"
    },
    {
      "@type": "FAQPage",
      "mainEntity": [{
        "@type": "Question",
        "name": "Does GA4 track ChatGPT traffic?",
        "acceptedAnswer": {"@type": "Answer", "text": "GA4 assigns recognized AI assistant referrals to its AI Assistant channel."}
      }]
    }
  ]
}

Keep dateModified honest, since it should match the visible update date from signal 7.

Signals with emerging evidence

The last two signals rest on early research or on adoption data rather than citation outcomes. One is worth adopting as a writing habit, and the other you can safely skip for now.

9. Original data, statistics and cited sources

The strongest controlled evidence for any content tactic comes from academic work, though in a simulated setting. The GEO paper presented at KDD 2024 tested content edits across a benchmark of 10,000 queries and found that adding citations, quotations from relevant sources and statistics raised source visibility by up to 40%. On Perplexity, adding quotations improved visibility by 22% over the baseline, while keyword stuffing performed 10% worse.

The caveat is the setup. The benchmark used a simulated generative engine with a fixed pool of sources, and live engines retrieve differently. The researchers tested what helps once you're in the candidate pool. They didn't test how to get there.

The habit is still cheap and useful. Publish numbers you've measured yourself, each with its population and period, and link to the sources behind every claim you didn't originate. A specific, sourced figure gives an engine something worth citing that a competitor's rewrite of common knowledge can't supply.

10. llms.txt and AI-specific files

llms.txt is a proposed markdown file at the root of a domain that points AI agents to the pages they need. Jeremy Howard proposed it in September 2024 and revised it in August 2026.

No major AI company has confirmed that its crawlers read llms.txt for search or citation. Google's AI optimization guide, linked in the schema section, says Google Search ignores these files and that keeping one neither helps nor harms visibility. OpenAI, Anthropic and Perplexity publish llms.txt files for their own developer documentation, which serves coding agents rather than answer engines.

Usage data matches. 28% of the 137,000 domains using Ahrefs Web Analytics publish an llms.txt file, and 97% of those files received zero requests in May 2026. Keep one if you run developer docs that coding agents consume, and otherwise spend the hour on signal 1.

Track AI visibility and citations

AI visibility needs three measurements, because no single one captures it. GA4 shows the clicks that arrive, a prompt log shows what engines say, and share metrics turn the log into trend lines leadership can read.

Start with GA4. In May 2026 Google added an AI Assistant default channel that assigns the medium "ai-assistant" when the referrer matches a recognized assistant.

Google's examples include ChatGPT, Gemini, Claude and Copilot, but Perplexity isn't named and Google hasn't published the full list. Clicks from AI Overviews and AI Mode land in Organic Search, and some AI visits arrive with no referrer and show up as Direct.

Build a custom channel group that catches both the native channel and the referrers it misses. Use two conditions joined by OR (Medium exactly matches ai-assistant, or Source matches this regex), and drag the channel above Referral, because GA4 assigns each session to the first rule it matches:

chatgpt\.com|chat\.openai\.com|perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|claude\.ai

Then run a manual prompt log. Write 20 prompts the way buyers type them, split across category discovery, competitor comparisons, problem-first questions and validation questions that name your brand. Record these fields for every run:

  • Date, engine, model or mode, and whether web search was on
  • The exact prompt text
  • Whether your brand was mentioned, and its position in any list
  • Whether your domain was cited, and which URL
  • Competitors named and third-party domains cited
  • Any factual errors in how the engine described you

Three metrics come out of that log. Brand mention rate is answers naming your brand divided by answers checked. Citation share is citations to your domain divided by all citations in the set. Share of voice is your brand's mentions divided by all brand mentions across tracked competitors. Vendors define the denominators differently, so fix yours and keep it stable, as our guides on AI visibility metrics and measuring AI share of voice explain.

Repeated runs of the same prompt return different answers. Recommendation lists rarely repeat exactly, while the rate at which a brand appears across many runs is far steadier. Run each prompt several times per check, report the result as a range, and compare four-week trends rather than single readings.

Move from manual tests to ongoing tracking

A spreadsheet works for the first month and breaks once the prompt set, competitor list and engine count grow, because every check means rerunning and re-tagging by hand. GrowthX built CheckThat, which continuously monitors ChatGPT, Claude, Perplexity, Google AI and more, so the prompt set and competitor list are tracked on a schedule instead of rerun by hand. Book a demo to see how the GrowthX platform runs this workflow in one workspace, if maintaining it across separate tools is the bottleneck.

Traditional SEO still matters in a zero-click world

Fewer searches send clicks, and that makes search fundamentals more valuable rather than less. 68.01% of U.S. Google searches ended without a click in January through April 2026, based on Similarweb's desktop and mobile-browser panel, which excludes the Google app.

That figure argues for keeping both channels. Keep the technical work, the keyword research and the ranking reports, and add citation and mention tracking beside them so you can see when a page ranks well but never gets quoted.

AI Overviews draw their supporting links from Google's index, so a page that isn't indexed and snippet-eligible can't appear as a supporting link there at all. Rankings for relevant sub-questions can put a page into more of the result sets engines retrieve when they fan out a query.

How to optimize for AI search step by step

Work through these steps in evidence order, because each one depends on the ones above it:

  1. Audit robots.txt and your CDN or firewall rules. Allow OAI-SearchBot, Claude-SearchBot and PerplexityBot, decide on GPTBot, ClaudeBot and Google-Extended, and confirm OpenAI's published IP ranges aren't blocked.
  2. Rewrite your highest-value pages answer-first. Put the direct answer in the first sentence under each heading, name the subject in every section, and split any section covering two sub-questions.
  3. Map the third-party pages engines cite for your category prompts. Pursue inclusion in those roundups, keep review profiles current, and earn mentions in publications your buyers read.
  4. Check indexing in Google Search Console and Bing Webmaster Tools. Target the fan-out sub-questions around each priority query, not only the head term.
  5. Build Content Clusters with hub and spoke pages linked in both directions. Add named authors with real credentials, and refresh the substance of priority pages before changing their dates.
  6. Add Organization, Article and FAQPage JSON-LD that matches visible text. Treat it as entity description, not a citation lever.
  7. Set up the GA4 custom channel group and start a 20-prompt log this week. Run each prompt several times per check, record mentions and citations by engine, and report four-week ranges to leadership.

Related articles

  1. What Is LLM Seeding and How Does It Get You Into AI Answers? Learn · 16 September 2026
  2. What Is LLM Optimization? Learn · 16 September 2026
  3. How to Rank in AI Overviews Learn · 11 September 2026
  4. How to get recommended by AI search engines Learn · 16 July 2026
  5. Unlinked Brand Mentions Are the Citation Lever in AI Search Learn · 14 July 2026
  6. What Answer Engine Optimization Is and How to Earn AI Citations Learn · 6 July 2026
  7. How to improve your brand's visibility in AI search Learn · 2 July 2026
  8. How AI Search Engines Select and Cite Sources Learn · 2 July 2026