SEO
AI SEO / GEO: make a website citable by artificial intelligence
2026 GEO method
Making a website citable by AI involves three separate checks: technical eligibility, observed page citations and faithful reuse of its information. This guide applies them to dated records and an open lab, stating the conditions and limitations of each observation.
- Eligibility Access, indexing and engine-specific controls.
- Citation Exact URL, observed answer, date and context.
- Fidelity Meaning, figures, units, scope and exclusions preserved.
October 9, 2026 revision: Claude/GPT comparison, C variant, missing-information diagnostic and “your text” tool. Source consultations, T0/day-30 records and earlier series keep their own dates.
Citable definition
An AI-citable website separates eligibility, observed citation and faithful reuse.
An AI-citable website provides information that can be accessed, verified and attributed. Assess it with three separate questions: is the page eligible under the engine’s rules, is its exact URL cited in an observed answer, and does the answer faithfully represent what the source supports? Evidence for one does not establish the others.
Editorial clarity helps readers understand and check a claim. Keep a price, duration or capability together with its units, scope, conditions, exclusions and dated source. This precaution does not guarantee that an assistant will select or cite the page.
Here, GEO adds observation of citations and their fidelity to the usual SEO checks. The historical records and open lab below distinguish observed evidence, simulation and unresolved questions.
To make a website citable by AI, first check access and the engine’s eligibility rules. Make each claim verifiable by keeping numbers together with their units, conditions, exclusions and dated sources. Then observe exact-URL citations and the fidelity of the resulting answers separately. An accessible page may never be cited; a cited page may be misrepresented. No editorial format guarantees these outcomes.
For the French query “rendre un site citable par les IA”, the August 27, 2026 baseline records edikka.com in first position on Google France and among Perplexity’s cited sources. The Google AI Mode response is marked incomplete in the export; it cannot establish an absence of citation.
Evidence boundary. Edikka self-measurement, not an independent audit: one French query and one pass per surface and date. A domain citation does not necessarily identify this article. These traces establish neither a stable rate, English-market performance nor a causal effect of the writing.
Confirmed for Google: ordinary SEO and snippet eligibility remain the foundation. Contradicted for Google: llms.txt and special AI structured data are not requirements for generative Search visibility. Supported by the reviewed research: a citation-oriented rewrite can improve passage use while degrading retrieval. Not proven: any editorial format, by itself, causes stable citations across platforms.
Make the observation reproducible
Record the exact query, date, language, market, account state, device and product surface before interpreting the result.
Do not collapse discovery into citation
Log retrieval, use, brand mention, linked citation, source position and referral traffic as separate outcomes.
Publish enough material to challenge the claim
Keep raw exports or captures, a manifest, limitations, timestamps and checksums so another reviewer can audit the observation.
What T0 and day 30 show
| Surface | August 27, 2026 · T0 | September 26, 2026 · day 30 |
|---|---|---|
| Google France | Position 1 | Position 1 |
| Perplexity | Domain cited; CSV does not establish the exact URL | Incomplete response: indeterminate |
| Google AI Mode | Incomplete response: indeterminate | Exact URL found in the trace; response completeness uncertain |
October 8 review: the day-30 CSV marks Google AI Mode as complete, while the capture still appears to be generating. The cited URL is observable; a final answer and its overall fidelity remain indeterminate. Historical files have not been retroactively changed.
Business example
A correct price can support the wrong decision when its conditions disappear.
Edikka’s pricing policy of August 20, 2026 places a targeted redesign at EUR 5,000–9,999 excluding VAT: a journey, a landing page or a small site with up to three templates and available content. Multilingual work and bespoke business integrations are excluded. This range is neither a market average nor an automatic quotation.
The number stays, the scope changes
“A complete bilingual redesign with CRM costs EUR 5,000–9,999 including VAT.” The number resembles the source, but the tax basis, offer and exclusions have changed.
The amount stays tied to the offer
“Edikka lists EUR 5,000–9,999 excluding VAT for a targeted redesign of up to three templates with available content, excluding multilingual work and bespoke integration. A bilingual CRM project needs a different scope.”
These examples were authored to explain the distinction; neither is an observed assistant response.
Lab findings
What the lab shows
Three authored cases in French and English: price, duration and hosting. A distributes conditions; B groups some sentences without changing their wording; C attaches all four conditions to key sentences. The same lexical retriever selects one three-sentence fragment, with full text as the control.
Attaching conditions to a sentence improves their availability in two cases, but still fails in the third. This table counts conditions actually supplied to the model, out of four expected conditions. It measures neither citation frequency nor answer fidelity.
| Case | A · distributed | B · grouped | C · within the sentence | Full-text control |
|---|---|---|---|---|
| Price | 2/4 | 3/4 | 4/4 | 4/4 |
| Duration | 1/4 | 3/4 | 4/4 | 4/4 |
| Hosting | 3/4 | 1/4 | 1/4 | 4/4 |
For hosting, the chosen fragment covers PHP but leaves static capabilities elsewhere, even with C. C also changes length and repetition; our splitter keeps each sentence whole. Joint condition transport is therefore partly mechanical. No universal GEO writing rule is established. All 72 sensitivity settings and the adverse result are published.
Claude and GPT: same fragments, different diagnostics
Both models received the same questions, sentences and diagnostic categories. Each series has 107 complete responses and one retained transport failure; those failures affect different inputs. The comparison below therefore uses 106 complete matched trials.
| Model | Missing conditions recognised | Supplied conditions declared missing | Exact quotes with rejected associations |
|---|---|---|---|
| GPT-6 Astra | 74/83 | 9/341 | 9/341 |
| Claude Sonnet 5.5 | 78/83 | 0/341 | 5/346 |
Absent and supplied conditions are defined by prespecified coding. All extracts in the table are exact substrings; some cited sentences are not associated with that category in the coding. Categories can overlap: a discrepancy is not proof of hallucination. An exact quote alone therefore does not certify a faithful restatement.
This test explicitly asks models to declare gaps. Three authored cases, two languages and one model per provider establish neither a general model ranking nor a universal GEO writing rule. Medium effort, tokenizers and output constraints differ across providers. The A/B/C retrieval table is unchanged: it describes what is supplied before generation.
Claude series, 9 October 2026: USD 0.718656 estimated before tax, plus USD 0.032428 reserved for the request without a response, within the USD 2 cap. The failed request was not retried; only unattempted trials were continued. Historical series, including the biased pilot, remain separate and unchanged. No web-citation uplift is established.
Conditions
What needs to travel with a number
A number remains useful when readers can recover the information that gives it meaning. Use this grid to write and check a restatement; it is not an engine-mandated format.
| Element | Question to resolve | Edikka example |
|---|---|---|
| Amount and unit | Which number, currency or duration? | EUR 5,000–9,999; 4–7 weeks. |
| Tax basis | Excluding tax, including tax or another stated basis? | Range excluding VAT. |
| Scope | Which offer and deliverables does the number cover? | Targeted redesign: a journey, a landing page or up to three templates. |
| Conditions | Which prerequisites or starting point change the meaning? | Content available; schedule defined in the signed proposal. |
| Exclusions | What does the number leave out? | Multilingual work and bespoke business integration excluded. |
| Date | When were the information and its source valid? | Pricing policy dated August 20, 2026; consultation date kept separate. |
| Source and limits | Who claims what, and with which limitation? | Edikka policy: neither a market average nor an automatic quote. |
Verification grid
One grid, three stages, twelve checks
For each page, record “verified”, “needs correction” or “indeterminate” and attach the evidence. The last column gives the status of the supporting claim in the register, not your page’s test result. Local methods without a corresponding entry are labelled as such.
Eligibility: can the page be used?
| Check | Expected evidence | Correction | Basis and status |
|---|---|---|---|
| Page access | HTTP response, robots.txt and access for the relevant crawler. | Fix unintended blocking, errors or authentication. | GEO-018 · Confirmed for OAI-SearchBot. |
| Google indexing and snippets | URL Inspection: indexing and snippet eligibility. | Resolve noindex, inconsistent canonicals and snippet directives. | GEO-017 · Confirmed within Google’s scope. |
| Visible, verifiable content | Main text, figures and conditions accessible on the page. | Make the information readable; do not reserve conditions for a special AI file. | GEO-011 · Requirement for a special file contradicted. |
| Consistent markup | Compare JSON-LD claims with visible content. | Fix discrepancies; do not present special AI schema as mandatory. | GEO-014 · Special-schema requirement contradicted. |
Citation: which page is actually cited?
| Check | Expected evidence | Correction | Basis and status |
|---|---|---|---|
| Reproducible context | Exact query, language, market, surface and timestamp. | Complete the record before comparing observations. | GEO-038 · Supported: measurement variability. |
| Finished response | Observed generation state, stable text and final screenshot. | Wait for completion; retain indeterminate status when evidence is missing. | Collection choice · outside the register. |
| Exact URL | Visible source links and the resolved article URL. | Separate brand mention, domain citation and exact-page citation. | Measurement choice · outside the register. |
| Repetitions and missing observations | Numerator, denominator, incomplete runs and conditions of each wave. | Retain failures; do not count incomplete answers as absent citations. | GEO-038 · Supported; incomplete-response handling specified here. |
Fidelity: does the answer support the same decision?
| Check | Expected evidence | Correction | Basis and status |
|---|---|---|---|
| Conditions of the number | Amount, unit, tax, scope, prerequisites, exclusions, date and source. | Restore conditions whose omission changes the decision. | GEO-036 · Supported: citation and claim support differ. |
| Information actually supplied | Provided fragments and available conditions before generation, when accessible. | Distinguish retrieval loss from omission in the answer. | GEO-005 · Supported; real retrieval may remain unobservable. |
| Recognised gaps | Compare truly absent conditions with declared unavailable conditions. | Make uncertainty explicit; a structured declaration does not certify free prose. | Local experimental measurement · outside the register. |
| Conclusion and source | Compare the conclusion’s scope with the cited evidence. | Correct extrapolation; retain indeterminate status for unverifiable claims. | GEO-036 · Supported; no overall fidelity guarantee. |
2026 mechanics
ChatGPT, Perplexity, AI Overviews, llms.txt: what you actually need to remember.
AI engines do not all work the same way. Some rely on a search index, some browse the web for specific queries, some show visible citations and some crawlers have distinct roles. The right reflex is to separate access, indexing, citation, training and measurement.
| Topic | Documentation reviewed on October 8, 2026 | SEO / GEO decision |
|---|---|---|
ChatGPT and OpenAI | OpenAI distinguishes OAI-SearchBot for Search discovery from GPTBot for potential model improvement and ChatGPT-User for user-triggered visits. Its publisher guidance says OAI-SearchBot must not be blocked if page content is to be included in summaries and snippets. | Treat crawler access as eligibility, not as a ranking or citation guarantee. Allow or block Search and training independently, then verify the live robots.txt and server response. |
Perplexity | Perplexity cites sources in answers and documents dedicated crawlers for different uses. | Make facts, conditions and sources explicit; then observe which URLs are actually cited and whether their claims are faithfully restated. |
Google AI Overviews and AI Mode | Google documents Preferred Sources in AI Overviews and AI Mode: a badge can identify sites selected by the user. This preference applies to a domain or subdomain and requires inclusion in Search generative AI features. | Optimize indexing and usefulness first. Treat Preferred Sources as a user-level domain preference, not as a general citation or ranking guarantee. |
llms.txt | Google says it does not use llms.txt or other special AI text files for Search; publishing one neither helps nor harms visibility or rankings in Google Search. | Maintain llms.txt only for services or editorial workflows that use it. Do not present it as a Google visibility lever or a substitute for the visible page. |
Structured data | Google says structured data is not required for generative AI Search and that no special AI schema is needed. Ordinary structured data can still support standard Search features when it matches visible content. | Use Article, Person, Organization, BreadcrumbList or FAQPage only when they describe the page accurately. Schema is metadata, not proof of citation readiness. |
Platform rules and editorial choices
Google’s guide reviewed on October 8, 2026 does not prescribe artificial chunking, an ideal length or a special AI format. The formats here are editorial tools. For generative Search features, also check indexing, snippet eligibility and the Search Console inclusion control; these settings do not guarantee citation.
Further GEO measurement
Further reading
These guides explain the instruments and settings used by this grid.
Sources
Official sources used to date the recommendations.
The platform documentation below was reviewed on October 8, 2026. This is our review date, not the sources’ publication date or evidence of a fresh crawl of Edikka. T0/day-30 observations, the pricing policy and the lab series retain their respective dates and versions. Check the relevant engine’s current rule again before changing access or indexing settings.
- OpenAI
OpenAI crawlers, roles of GPTBot, OAI-SearchBot and ChatGPT-User
- Perplexity
- Google
Google guide to optimizing a site for generative experiences
- Schema.org
ItemList: describe the seven conditions in the visible table
- llms.txt
llms.txt proposal for exposing a content map to language models
Conclusion
Check the page, then verify what answers do with it.
Start with one priority page: verify access and sources, and keep each figure with its conditions. Record the query, engine, date, cited URL and reused passage. No citation, an incomplete answer and a distorted claim are different observations; choose a correction that addresses the defect actually found.
Prompt, full source and citation
Explore this analysis.
Choose a task, review the prompt and take it to your AI.
AI SEO / GEO: make a website citable by artificial intelligenceUpdated October 9, 2026
Read the full prompt
Your inputs stay in this browser until you paste them into your AI. No analysis runs here.
Your AI cannot read the link?
Copy the full source or download the document, then attach it to your conversation.
Full message to copy
No message is sent from this page.
Review and publish the post yourself on LinkedIn.
Information should remain accurate when it leaves your website.
AI can amplify both an idea and its errors. A price without its scope, a timeline without its starting point or a capability presented as certain can steer someone towards the wrong decision.
At Edikka, we work to make every claim verifiable, including the conditions that give it meaning. Our lab showed that grouping information can also leave useful conditions outside the selected fragment. That counterexample matters: it requires us to test our advice and publish its limits.
Our standard: measure what reaches the reader, retain adverse findings and revise the method when the evidence calls for it.
Frequently asked questions
Additional answers to clarify the key points covered in this article.
What is an AI-citable page?
An AI-citable page provides information that is accessible, verifiable and attributable. Distinguish eligibility under the engine’s rules, observed citation of the exact URL and faithful reuse of the source. Numbers should remain tied to their units, conditions, exclusions and sources. None of these qualities alone guarantees a citation.
What is the difference between classic SEO and GEO?
Classic SEO mainly targets crawl, indexation, ranking and clicks from search results. GEO adds an answer logic: the page must be selectable, summarizable, citable or usable as a source by an answer engine. SEO fundamentals remain necessary, but they are not always enough to make a passage usable by AI.
Which conditions should accompany a number restated by AI?
Keep the amount and unit, tax basis, scope, prerequisites, exclusions, date and source. Then check what reaches the selected fragment and what remains in the answer. The lab shows that a more complete sentence does not guarantee that every condition will be retrieved.
Are structured data enough to be cited by AI?
No. Markup describes the page and must match visible content. Google requires neither structured data nor special schema for its generative features. Correct JSON-LD does not establish page selection, an exact URL citation or fidelity of the restatement.
Should you create an llms.txt file to be cited?
Google says it does not use llms.txt for Search and that the file has no effect on visibility or rankings. It may serve an identified service or editorial workflow, but it does not replace the visible page or guarantee a citation.
Should you allow GPTBot, OAI-SearchBot and AI crawlers?
The decision depends on the site strategy. You need to distinguish search crawlers, user-triggered crawlers and crawlers used for model improvement. Allowing or blocking a user-agent should be an explicit choice, documented in robots.txt and aligned with the goal: visibility, content protection or controlled exposure.
How can you check whether an AI restates a price accurately?
Compare the answer with the dated source: amount and currency, tax basis, offer, scope, prerequisites and exclusions. Keep the exact question, answer and cited URL. Where supplied fragments are known, distinguish a condition missing at retrieval from one supplied but omitted or distorted in the answer. An exact extract cannot certify the whole answer; a point that cannot be verified remains indeterminate.
How do you measure whether a page is cited by AI?
Keep the exact query, language, market, surface, date, response and source URLs. Separate a domain mention, an article citation and support for each claim. An incomplete response remains indeterminate. Separate supplied text, a supplied URL and search without a URL; API outputs do not automatically represent consumer interfaces.
Do AI Overviews change how a page should become citable?
Yes, because Google AI search can synthesize several sources and expose supporting links in a more conversational context. But the basics remain the same: useful, indexable content, accessible snippets, clear structure and enough reliability to be chosen as a source.
Where should you start to make a page more citable?
Choose one sensitive fact on a useful page: a price, duration or capability. Keep the fact close to its source, scope and exclusions, without imposing a format. Then compare restatements under a frozen protocol. This page’s demo illustrates the checks; its authored examples are not a performance measurement.
