Quick answer

There's no button or markup that guarantees a citation in ChatGPT, Perplexity, Gemini, or Claude. To raise a measurable probability, first make sure the right crawler has access, publish a page that provides evidence attributable to a specific question, then test a prompt panel with several runs. Count brand mentions and cited URLs separately.

Controls differ by provider. OpenAI, for instance, distinguishes OAI-SearchBot for Search from GPTBot for training; Perplexity distinguishes PerplexityBot from its user-triggered agent; Anthropic documents Claude-SearchBot, Claude-User, and ClaudeBot. Blocking or allowing them is an access decision, not a ranking optimization. For Gemini and Google, ordinary Search rules remain the foundation.

Key takeaways

  • Allowing a bot makes retrieval possible; it guarantees neither indexing, citation, nor favorable tone.
  • Search, training, and user-query bots can be distinct: don't copy a robots.txt block without understanding its effect.
  • A credible experiment fixes prompts, market, language, date, platform, and repeats before observing results.
  • The citation rate uses every run in the denominator, including answers with no brand.
  • Citations vary; publish the distribution and the failures instead of a cherry-picked screenshot.

Step zero: distinguishing access, selection, and evidence

Control What it does What it doesn't do
robots.txt / WAF Allows or blocks certain declared agents Doesn't ask the engine to recommend the brand
Search indexing Makes a page eligible for Google surfaces per its rules Doesn't guarantee AI Overview or AI Mode
Content with sources Provides a verifiable unit of information Doesn't force attribution
Structured data Describes the visible content with a vocabulary Isn't an "AI citation" schema
Prompt panel Measures a presence observed on a date Doesn't represent all market demand
Analytics / CRM Observes attributable visits and conversions Doesn't capture every influence with no click

Most GEO promises confuse these lines. An audit that finds "bot allowed" can't conclude "you'll be cited." A tracker that finds a citation can't conclude "it generated revenue."

Controlling agents without mixing up search and training

OpenAI

OpenAI's official documentation states that OAI-SearchBot is used to make sites appear in ChatGPT's search features. An exclusion stops Search answers from showing the site, except possible navigation links. GPTBot concerns content that may be used for training; its rules are independent. ChatGPT-User executes certain actions requested by a user and isn't the Search control mechanism (OpenAI, Overview of OpenAI crawlers). The page states that a robots.txt change can take about 24 hours to be picked up. This is a current product description, not a citation guarantee.

Perplexity

Perplexity explains that PerplexityBot is used to surface and link sites in its search results, while Perplexity-User can fetch a page at a user's request and may ignore robots.txt. The provider publishes IP ranges and recommends checking both the agent and the IP in a WAF (Perplexity Crawlers documentation). The ranges change: don't freeze a copied list in an article.

Claude

Anthropic distinguishes Claude-SearchBot for improving Search results, Claude-User for user-initiated queries, and ClaudeBot for collection that may contribute to training. Its help center article from 7 April 2026 states that blocking SearchBot or User can reduce visibility in the corresponding search (Claude Help Center). Again, access doesn't mean selection.

Google and Gemini

For Google Search's generative features, Google states no additional technical requirement exists: a page must be indexed and eligible for a snippet. Documented Search controls apply (Google, AI features and your website). Don't automatically transpose this documentation to the Gemini app outside Search; surfaces and measurement can differ.

Seven-step protocol

1. Define the scope before the outcome

Fix country, language, audience, offer, platforms, and period. Example: France, French, SME SaaS executives, ChatGPT Search, Perplexity, and Gemini, over four weeks. A visible model change creates a new period.

2. Build a stratified panel

Create at least four strata: discovery, problem, comparison, and decision. A panel of 100 prompts can contain 25 questions per stratum. Prompts come from sales, support, Search Console, and communities; a Reddit question proves a qualitative need, not its volume.

3. Freeze the text and the rules

Assign an ID to each prompt. Don't fix a question after an unfavorable answer. Fix the number of runs (e.g. five per prompt and per platform) and the deduplication rules before launching the test.

4. Verify technical access

Check robots.txt, the WAF, HTTP responses, the canonical page, and the render. Verify IP ranges from the official URL at the time of the test. Don't blindly allow a user agent on the text string alone; it can be spoofed.

5. Publish evidence worth attribution

For each intent, provide a data point, a method, an example, or a tool that the generic synthesis doesn't have. Name the author, date, assumptions, and limitations. An extraordinary claim with no evidence isn't "extractable"; it's just easy to copy.

6. Execute and code the results

For each run, record: the answer obtained, brand mention, domain citation, exact URL, citation position, context, competitors, and error/absence. Keep the raw results when terms of use allow, and don't collect any unnecessary personal data.

7. Connect to behavior

Add identifiable referrals, conversions, brand searches, and business feedback, without summing them into an arbitrary score. A simultaneous rise after publication is an association. A well-documented cohort test or interrupted series strengthens the inference without making it perfect.

The right formulas

KPI Formula Denominator to show
Mention rate runs containing the brand / valid runs All prompts × repeats × platforms, minus errors defined before the test
Citation rate runs containing a domain URL / valid runs Same base as the mention rate
Citation share of voice domain's citations / citations of all tracked domains Define whether a run can count multiple times
Prompt coverage prompts with at least one citation / prompts tested Unique prompts, not runs
Volatility prompts whose status changes between runs / repeated prompts Same text, same period, same platform
Referral conversion attributed conversions / identified referral sessions An explicit attribution rule

Worked example

An agency tests 60 prompts on three platforms, four times each: 60 × 3 × 4 = 720 runs. Twenty runs fail per an error rule defined before launch; 700 remain valid. The brand appears in 98 answers and its domain in 42.

  • mention rate: 98 / 700 = 14%;
  • citation rate: 42 / 700 = 6%;
  • mention-citation gap: 8 points;
  • if 17 of the 60 prompts get at least one citation, coverage: 17 / 60 = 28.3%.

The report should show the distribution by stratum and platform. A global rate can hide 20% on brand queries and 0% on non-brand comparisons. The protocole-visibilite-ia.csv sheet provides the needed columns.

What external data proves and doesn't prove

An Ahrefs analysis updated in March 2026 examined 863,000 SERPs and four million URLs cited in AI Overviews. It found 37.9% of cited URLs in Google's top 10 blocks for the same query, 31.2% in blocks 11 to 100, and 31.0% beyond (Ahrefs). Geography, device, exact dates, and keyword selection aren't fully disclosed; the parser has changed. The study shows an imperfect overlap, not that "62% of citations ignore Google" nor that fan-out is the sole cause.

The GEO research paper published at KDD 2024 tested 10,000 queries with a synthetic generative engine based on GPT-3.5 and observed that citations, statistics, and credible quotations improved its visibility metrics, while keyword stuffing didn't help (GEO paper DOI). The models and setup are historical; it isn't a 2026 recipe for commercial platforms. The robust lesson is to test proven content, not to guarantee an effect.

Mistakes and stopping conditions

  • Testing only once. Stop the interpretation if no repeat measures volatility.
  • Changing prompts after seeing the answers. Create a new version of the panel.
  • Comparing platforms with different denominators. Recalculate on a common panel or show them separately.
  • Allowing every bot with no governance decision. Search and training are separate choices.
  • Blocking by user agent with no IP check. A string can be spoofed; follow the provider's WAF guidance.
  • Fabricating messages on Reddit or Quora. Astroturfing creates reputational risk and proves no causality.
  • Declaring a win on branded prompts. Separate branded, non-branded, comparison, and decision.

How SEOryon fits in

SEOryon tracks mentions/citations on ChatGPT, Perplexity, Gemini, and Claude and can connect that observation to research, production, and content publishing. Its public score checks a URL on 27 signals. The team still needs to define the panel, validate bots/WAF, review a mention's context, and connect observations to business data. A tracker reduces the work; it doesn't turn a correlation into causation.

Measurable exercise

Create a panel of 40 prompts in four strata, then run three runs on two platforms. The deliverable must contain 240 expected rows, an error log, rates with denominators, the distribution by stratum, and five analyzed unfavorable answers. Success: a second analyst reproduces the three main rates from the raw file.

Where to go next

FAQ

Is allowing OAI-SearchBot enough to appear in ChatGPT?

No. It allows Search access per OpenAI's documentation. Selection depends on the query and OpenAI's systems; no citation is guaranteed.

OpenAI documents separate controls for GPTBot and OAI-SearchBot. Anthropic also distinguishes its agents. Check each current policy and your legal decision before changing anything.

Should you allow Perplexity-User in robots.txt?

Perplexity states this user-triggered agent can ignore robots rules. Consult its documentation and manage the WAF using the published IP ranges, rather than assuming behavior identical to the automatic crawler.

Does structured data get a page cited?

No official source guarantees this. It can make content explicit and support compatible features, provided it matches what's visible. Test the effect instead of promising it.

How many runs do you need?

There's no universal threshold. Start with three to five repeats per prompt and per platform, then increase if volatility is high or the financial decision justifies it. Always publish the number.

References

Method note

Method: provider pages verified 16 July 2026, translated and edited 22 July 2026. User agents, IP ranges, and behaviors can change; revalidate their official pages right before any robots.txt or WAF change.