# AI Search Ranking Factors 2026: 12 Signals Ranked by Evidence

> A source-backed audit of 12 AI search ranking signals across Google AI, ChatGPT, Claude, and Perplexity—separating confirmed mechanics from research and hype.

Author: Vincent Ruan (https://x.com/0xVinceAI)
Publisher: Attrifast — https://attrifast.com
Last updated: 2026-08-22
Canonical: https://attrifast.com/blog/ai-search-ranking-factors-2026

**TL;DR**

- There is no vendor-published universal list of “AI ranking factors.” This audit separates **five platform-confirmed mechanisms**, **three research-supported content patterns**, and **four widely repeated but unproven shortcuts**.
- The durable work is unglamorous: allow the correct crawler, make the page indexable, match the query, publish original evidence, and link the page from relevant hubs.
- The KDD 2024 GEO study tested 10,000 queries. Citations, quotations, and statistics improved visibility in its benchmark, with gains of up to 40% across queries; that is not a guaranteed production lift.
- Schema, `llms.txt`, a fixed FAQ count, and a fixed number of `sameAs` links may be useful for other reasons, but no major engine documents them as direct AI citation boosts.
- Measure citations, visits, conversions, and revenue separately. A higher citation count is not the same as customer acquisition.

AI search ranking factors are best understood as an **evidence ladder**, not a secret checklist. Google, OpenAI, Anthropic, and Perplexity document how publishers can make content accessible to their search products. They do not publish production scoring formulas, fixed weights, or a universal recipe for getting cited.

This review evaluates 12 commonly recommended signals against vendor documentation and controlled research. It deliberately avoids turning correlation—such as “high-authority domains are cited often”—into a claim about an undisclosed algorithm.

## The 12 AI search signals at a glance

| # | Signal or tactic | Evidence grade | What the evidence supports | What it does **not** prove |
|---:|---|---|---|---|
| 1 | Crawl and index eligibility | **Confirmed** | Google AI Search links must be indexed and snippet-eligible | Indexing guarantees citation |
| 2 | Correct AI search crawler access | **Confirmed** | OpenAI, Anthropic, and Perplexity publish separate search crawler controls | Allowing a bot guarantees ranking |
| 3 | Query and passage relevance | **Confirmed mechanism** | Search-grounded systems retrieve content relevant to the request | A published keyword-density formula |
| 4 | Useful, original information | **Confirmed guidance** | Google explicitly prioritizes unique, non-commodity, people-first content | “Original” content always outranks stronger evidence |
| 5 | Internal discovery and anchor context | **Confirmed for Google Search** | Crawlable descriptive links help Google discover and understand pages | The same weight applies to every AI engine |
| 6 | Verifiable citations | **Research-supported** | The GEO study found citation addition improved benchmark visibility | A fixed production uplift on every engine |
| 7 | Relevant quotations | **Research-supported** | Quotations improved visibility in the GEO benchmark | More quotations are always better |
| 8 | Relevant statistics | **Research-supported** | Statistics improved visibility in the GEO benchmark | Unsupported numbers or decorative charts help |
| 9 | Structured data as a special AI boost | **Unproven** | Accurate schema aids semantics and supported Search features | JSON-LD is a direct AI citation factor |
| 10 | `llms.txt` as a ranking signal | **Unproven** | It is a community proposal and low-cost experiment | Google uses it; Google says it ignores the file |
| 11 | A fixed FAQ or heading count | **Unproven** | Clear answers improve readability and retrieval fit | Four FAQs or question H2s unlock citations |
| 12 | A fixed entity or `sameAs` count | **Unproven** | Consistent identity information reduces ambiguity | Four profiles or Wikidata guarantee recognition |

The classification is conservative by design. “Confirmed” means the platform documents the mechanism or eligibility requirement—not that it publishes a weight.

## How AI search retrieval changes the SEO question

Classic SEO asks, “Which page ranks for this query?” AI search adds two more stages:

1. **Retrieval:** which documents or passages enter the candidate set?
2. **Synthesis:** which retrieved evidence is used, summarized, or cited in the answer?

Google documents retrieval-augmented generation and query fan-out for AI Search. OpenAI documents a search-specific crawler, but not its scoring model. Anthropic and Perplexity likewise publish crawler controls without revealing factor weights.

The result is a simple operating rule: optimize first for **eligibility and evidence**, then test presentation. A perfectly formatted answer cannot be cited if the system cannot discover it; a crawlable page still will not be selected if it adds nothing useful.

## Confirmed signal 1: crawl and index eligibility

For Google AI Overviews and AI Mode, Google says a page must be indexed and eligible to appear with a snippet. Normal Search controls still apply, including `noindex`, `nosnippet`, `max-snippet`, and crawl permissions.

Check the basics before rewriting content:

- The canonical URL returns `200` and is not blocked by robots.txt.
- The rendered HTML contains the core answer.
- The page has a self-referencing canonical.
- Important images have descriptive alt text, but the critical evidence also exists as text.
- The page is linked from at least one relevant, crawlable page.
- The sitemap and internal links point to the canonical version.

Eligibility is binary; citation is competitive. Passing the first does not guarantee the second.

## Confirmed signal 2: allow the correct search crawler

Training crawlers and search crawlers are not interchangeable.

| Platform | Search or answer-retrieval token | Other documented token | Practical implication |
|---|---|---|---|
| OpenAI | `OAI-SearchBot` | `GPTBot` for model training; `ChatGPT-User` for user-triggered visits | A publisher can allow Search visibility while separately controlling training |
| Anthropic | `Claude-SearchBot` | `ClaudeBot` for training; `Claude-User` for user requests | Treat each control independently |
| Perplexity | `PerplexityBot`; `Perplexity-User` for user requests | Official IP ranges are published for verification | Do not rely on user-agent text alone at the WAF |
| Google | `Googlebot` for Search | `Google-Extended` controls certain Gemini uses, not Search ranking | Blocking Google-Extended does not remove a page from Google Search |

The official [OpenAI](https://developers.openai.com/api/docs/bots), [Anthropic](https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler), and [Perplexity](https://docs.perplexity.ai/docs/resources/perplexity-crawlers) documents are the source of truth. CDN or WAF rules can block a crawler even when robots.txt allows it, so verify request logs when inclusion matters.

## Confirmed signal 3: query and passage relevance

Retrieval systems need a passage that answers the actual request. This does not mean repeating an exact-match keyword. It means making the relationship between question, answer, and evidence easy to resolve.

A strong answer block usually contains:

- a direct answer in the first paragraph;
- the conditions under which the answer changes;
- a number, example, or primary source where appropriate;
- a boundary that prevents overgeneralization;
- descriptive headings that reflect real subquestions.

For this article, the primary query is **AI search ranking factors**. Secondary queries—ChatGPT ranking factors, AI citation factors, Perplexity ranking factors—are answered in engine-specific sections rather than repeated unnaturally in every paragraph.

## Confirmed signal 4: useful, original information

Google's 2026 guidance says unique, non-commodity content and first-hand experience become more important as AI can reproduce generic summaries cheaply. Originality can take several forms:

| Type of original value | Strong example | Weak imitation |
|---|---|---|
| Primary data | Published sample, date range, definitions, and limitations | A percentage with no denominator |
| First-hand test | Reproducible method and screenshots | “We tested it” with no method |
| Expert synthesis | Conflicting sources reconciled and graded | Ten sources summarized without judgment |
| Operational artifact | Checklist, query set, template, or code | A generic list of tips |
| Correction | Explains what popular advice gets wrong | Repeats the consensus to look safe |

Originality is not an excuse to invent precision. If a number cannot be audited, omit it or label it as an illustrative example.

## Confirmed signal 5: internal discovery and anchor context

Google's [link guidance](https://developers.google.com/search/docs/crawling-indexing/links-crawlable) says crawlable links and descriptive anchor text help Google discover pages and understand what they are about. This is directly relevant to Google AI because its Search experiences are grounded in the Search index.

Build a topic path instead of a footer farm:

- A hub should link to the strongest research pages.
- Each research page should link to its prerequisite and next action.
- Product pages should link to evidence that substantiates their claims.
- Blog CTAs should link to the product outcome, not only to a signup form.

This article sits between [where Google AI gets its information](/blog/where-does-google-ai-get-its-information) and the practical guide on [how to get cited by AI engines](/blog/how-to-get-cited-by-ai-engines). The broader taxonomy lives in the [AI Search Hub](/ai-search).

## Research-supported signals 6–8: citations, quotations, and statistics

The KDD 2024 paper [Generative Engines and Search](https://arxiv.org/abs/2311.09735) introduced GEO-bench with 10,000 queries. In its controlled experiments, adding citations, quotations, and statistics could improve source visibility; the paper reports gains of up to 40% across queries and up to 37% on Perplexity in its evaluated setting.

Those are benchmark results, not a promise that adding a chart will increase production citations by 40%. The responsible interpretation is narrower:

1. Claims that can be verified are easier to use safely.
2. Quotations can preserve authoritative wording when paraphrase would change meaning.
3. Relevant statistics make an answer more specific and less interchangeable.
4. Every number still needs a source, denominator, date, and boundary.

The wrong interpretation is to decorate a page with unsourced numbers. That weakens E-E-A-T and gives an AI system more claims it should avoid.

## Unproven shortcut 9: schema as a special AI ranking boost

Use schema where it accurately represents visible content. It can support Search understanding and eligibility for documented rich results. But Google's [structured-data policies](https://developers.google.com/search/docs/appearance/structured-data/sd-policies) do not promise ranking, and Google says no special AI markup is required for AI Overviews or AI Mode.

FAQ markup deserves extra caution. Google sharply limited FAQ rich results and deprecated HowTo rich results in 2023. A visible FAQ can still improve the article; the markup is not evidence of an AI ranking boost.

## Unproven shortcut 10: llms.txt as a ranking signal

`llms.txt` is a community proposal, not a substitute for web standards. Google's [2026 AI optimization guide](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide) says Google Search ignores it. OpenAI, Anthropic, and Perplexity do not publicly list it as a ranking signal.

Maintaining the file may still be reasonable when the cost is low and the content is accurate. The tradeoff is opportunity cost: fix canonical tags, crawl blocks, HTML content, sitemaps, and internal links first.

## Unproven shortcut 11: a fixed FAQ or heading count

Question-shaped headings can improve usability when they reflect real questions. A concise answer can create a useful retrieval passage. Neither observation proves that an engine requires four FAQs, an 80-word answer, or a heading every 300 words.

Write the number of sections the subject needs. This page uses five FAQs because they cover the recurring decision points; not because five is a ranking threshold.

## Unproven shortcut 12: fixed entity or sameAs counts

Consistent entity information is good publishing hygiene. Organization and Person schema can clarify identity when the markup matches the visible page. No public vendor document says that four `sameAs` profiles, a Wikidata entry, or a specific number of brand mentions produces an AI citation lift.

Avoid artificial directory and mention campaigns. They create inconsistent facts and weak pages—the opposite of the reliable identity you are trying to establish.

## How the priorities differ by engine

| Engine or surface | Publicly documented control | Reliable conclusion | Unknown |
|---|---|---|---|
| Google AI Overviews / AI Mode | Googlebot, Search indexing and snippet controls | Core SEO and people-first content remain relevant | Individual factor weights and citation selection |
| ChatGPT search | OAI-SearchBot | Search inclusion can be controlled separately from GPTBot training | Ranking formula, passage count, domain thresholds |
| Claude web search | Claude-SearchBot | Search access can be controlled separately from ClaudeBot | Ranking formula and citation weighting |
| Perplexity | PerplexityBot and Perplexity-User | Bot and user-request access are documented | Exact ranking and freshness weights |

Platform differences are real, but most tactical lists exaggerate what is known. Start with a shared technical and editorial foundation; then run per-engine tests against a stable prompt set.

## A 30-day implementation and measurement plan

| Week | Work | Output | Success check |
|---:|---|---|---|
| 1 | Technical eligibility audit | Crawl, index, canonical, bot, WAF, and HTML checklist | Target pages retrievable by intended search crawlers |
| 2 | Evidence upgrade | Direct answers, primary citations, methods, tables, limitations | Every material claim traceable to a source or method |
| 3 | Topic graph | Hub links, contextual links, descriptive anchors | Every priority page has relevant inbound and outbound links |
| 4 | Controlled measurement | Fixed prompt set, baseline, annotated publish date | Citation, session, conversion, and revenue trends reported separately |

Do not change all variables at once if you want to learn what worked. Update a defined set of pages, record the date, and compare against a control group or at least a stable baseline.

## Research standard and limitations

This audit was re-reviewed on August 22, 2026. **Platform-confirmed** means a vendor documents the mechanism or control. **Research-supported** means a published study measured an effect in its test environment. **Unproven** means the tactic may be useful but lacks public evidence as a direct ranking signal.

No outside observer can see the complete production ranking stack for ChatGPT, Claude, Perplexity, or Google AI. Citation results vary by prompt wording, location, account state, freshness, and product version. This page therefore ranks the strength of evidence—not secret algorithm weights.

## FAQ

### What are the most important AI search ranking factors in 2026?

The strongest public evidence supports five fundamentals: crawl and index eligibility, access for the relevant search crawler, relevance to the user's question, useful original content, and internal discovery through descriptive links. Controlled GEO research also supports adding verifiable citations, quotations, and statistics. Vendors do not publish a universal list or factor weights.

### How does ChatGPT decide which pages to cite?

OpenAI confirms that OAI-SearchBot is used to surface websites in ChatGPT search, but it does not publish the ranking formula or factor weights. A page must be discoverable and useful to the query; beyond that, claims about fixed domain-authority thresholds, schema boosts, citation counts, or passage limits are hypotheses unless OpenAI documents them.

### Does schema markup improve AI search rankings?

Schema can help search engines understand content and qualify a page for supported rich results when it matches visible content. There is no public evidence that adding JSON-LD creates a special ranking boost in ChatGPT, Claude, Perplexity, or Google AI. Use accurate schema for semantics and eligibility, not as a guaranteed AI citation tactic.

### Does llms.txt help a page rank in Google AI or ChatGPT?

Google explicitly says Google Search ignores llms.txt. The file remains a community proposal, and OpenAI, Anthropic, and Perplexity do not publicly document it as a ranking signal. It may be a low-cost experiment for machine-readable guidance, but it should never replace crawlable HTML, robots.txt controls, sitemaps, or internal links.

### How should I measure whether AI search optimization worked?

Measure the full funnel: prompt-level citation or mention rate, recognized AI referral sessions, conversion rate, and verified revenue by engine. Citation growth without clicks is visibility; clicks without conversions are traffic; only session-to-payment joins show commercial impact. Keep a before-and-after prompt set and annotate every material content change.

## Related research and next actions

- [Where Google AI gets its information](/blog/where-does-google-ai-get-its-information)
- [How to get cited by AI engines](/blog/how-to-get-cited-by-ai-engines)
- [AI Search Hub: 36 guides](/ai-search)
- [AI citation tracker: citation to revenue](/features/ai-citation-tracking)
- [AI referral traffic analytics](/ai-revenue-attribution)
- [Track website traffic from first click to revenue](/track-website-traffic)

## Sources

1. [AI features and your website](https://developers.google.com/search/docs/appearance/ai-features) — Google Search Central
2. [Top ways to ensure your content performs well in Google's AI experiences on Search](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide) — Google Search Central
3. [Creating helpful, reliable, people-first content](https://developers.google.com/search/docs/fundamentals/creating-helpful-content) — Google Search Central
4. [OpenAI crawlers and user agents](https://developers.openai.com/api/docs/bots) — OpenAI
5. [Anthropic web crawler controls](https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler) — Anthropic
6. [Perplexity crawlers](https://docs.perplexity.ai/docs/resources/perplexity-crawlers) — Perplexity
7. [Generative Engines and Search](https://arxiv.org/abs/2311.09735) — KDD 2024 / arXiv
8. [General structured data guidelines](https://developers.google.com/search/docs/appearance/structured-data/sd-policies) — Google Search Central
9. [Changes to HowTo and FAQ rich results](https://developers.google.com/search/blog/2023/08/howto-faq-changes) — Google Search Central