How websites reach AI answers — and what the evidence shows in 2026

Separate five different outcomes first

“Appearing in AI search” hides several stages. A crawler must fetch the page, a search system must index it, an answer system must retrieve it for a particular question, the model must use or cite its evidence, and a person may then visit the site and complete a useful action. Success at one stage does not prove success at the next.

A critical survey of 45 studies, released as a preprint in July 2026, describes the same multi-stage problem: discovery, retrieval, citation, influence on the answer, and economic outcome require separate measurements. Its author found no reviewed method that demonstrated stable, causal improvements to organic discoverability and downstream behaviour across platforms over time.

What the GEO experiment did—and did not—show

The peer-reviewed KDD 2024 paper by Aggarwal and colleagues introduced GEO, or generative engine optimization, and a benchmark of 10,000 queries. In its experiments, already-selected source documents were revised with relevant citations to reliable material, verifiable numbers, expert quotations, and clearer prose. The best variants increased the authors’ source-visibility metrics by up to 40%, with material variation by domain. In one Perplexity experiment, keyword stuffing performed worse than the baseline.

That headline number is not a promise of “40% more mentions.” The source was already among the documents supplied to the generative system; the study did not test whether an unknown page would be discovered more often in open search, improve conventional rankings, earn clicks, or generate leads. The 2026 critical survey identifies this as a central limitation of popular GEO claims.

• Use a number only with its method, date, unit, sample and primary source.
• Quote an expert only when their name, role, context and original recording or publication are known.
• Improve clarity without splitting prose into artificial fragments for an assumed algorithm.
• Do not repeat a query throughout the page: the experiment did not support keyword stuffing.

Make the page technically eligible

Google has no separate technical scheme for AI visibility. A page must be accessible to Googlebot, indexed, and eligible to show a conventional Search snippet. Important content should exist as text, internal links should make it discoverable without site search, structured data should match visible content, and the canonical URL should return a successful response without a redirect chain.

Other systems may use their own crawlers. ChatGPT search discovery depends on access for OAI-SearchBot, including through the CDN or web application firewall. Perplexity recommends allowing PerplexityBot and checking its published IP ranges. A general `User-agent: *` allowance works only when later rules, firewalls and bot protection also permit the request.

• Use one self-contained canonical URL returning HTTP 200 for each distinct reader task.
• Avoid `noindex`, crawler blocks, mandatory login, CAPTCHA or JavaScript challenges before content can be read.
• Expose the main answer and evidence in source or correctly rendered HTML.
• Link the page from navigation and a current sitemap; submit important changes through supported tools.
• Separate conventional search crawlers, OAI-SearchBot, PerplexityBot and user-triggered fetchers in server logs.

Publish material an answer can use precisely

Each page should solve one durable customer task. State the answer early, then explain conditions, alternatives, procedure, limitations and evidence. A peer-reviewed 2026 study of five generative systems from Google, OpenAI and Perplexity found substantial differences among platforms and changes across runs and time. Optimizing copy around one observed answer is therefore a weak strategy.

A 2026 preprint analysing 602 controlled prompts associates deeper answer use with structure, topical alignment and extractable definitions, numbers, comparisons and procedures. This is descriptive evidence, not a causal ranking formula. The restrained practical lesson is to publish original information that can be verified and extracted without losing its conditions, instead of paraphrasing common knowledge.

• Define a term before expanding it, and name the actor responsible for each action.
• Attach period, sample, unit, method and original data to a number.
• Compare options under the same criteria, conditions and observation date.
• Give a procedure inputs, steps, exceptions and an acceptance test.
• Identify the author or organisation, review date and factual owner.
• Correct a durable page at its existing URL when the reader task has not changed.

Respect the contracts of Google, Alice AI, Bing, ChatGPT and Perplexity

Platforms expose different parts of the process. Google connects generative answers to its core Search index and explicitly says special AI files, artificial content chunking and separate “GEO tricks” are unnecessary. Yandex says Alice AI normally relies on relevant, informative, original pages that rank highly in Search. ChatGPT and Perplexity document separate search crawlers.

These contracts create eligibility, not placement. Do not merge measurements from different systems into one invented score: the same user question can generate different search fan-outs, select different pages, or trigger no external search in a particular answer.

Measure repeated runs, not one screenshot

The ACL 2026 study found generative outputs unstable across executions and dates. One response is therefore not a ranking position. Build a small set of genuine buyer questions; record platform, language, region, login state, date and exact wording; and repeat observations with an unchanged protocol.

Count page access, source retrieval, visible citation, claim correctness, referral and business action separately. Bing warns that citation count does not reveal placement, authority or a page’s role in an answer. Likewise, an unlinked brand mention cannot automatically be attributed to work on one page.

1. Select a manageable set of questions from sales, support and conventional search data; separate informational and purchase tasks.
2. Save several natural paraphrases of each question without changing its intent.
3. Establish a baseline through repeated runs on every important platform.
4. Release one substantive content change and wait for confirmed recrawling.
5. Repeat the protocol and manually check whether each citation supports the adjacent claim.
6. Compare changes by platform and page; expand only effects that persist across repeated measurements.

A citation is not proof of faithful reproduction

A peer-reviewed EMNLP 2023 audit examined four generative search engines available at the time. On average, citations fully supported only 51.5% of generated sentences, while 74.5% of citations supported their associated sentence. These percentages are not estimates of 2026 products, but they establish a separate risk: a system can cite a page while distorting its meaning.

Visibility monitoring must therefore include correctness. Check how the system spells the company name, attributes prices and outcomes, separates service versions, and links to the source. A favourable but false mention is not success; it creates the wrong expectation in a customer’s decision.

• Maintain a primary page with unambiguous conditions, dates and scope.
• Do not publish a number without the context that determines its meaning.
• Track incorrect answers for the most important branded and product questions.
• Correct the source and connected pages; do not try to control the model with hidden instructions.

Tactics without reliable support

Google’s current guide specifically discourages artificial content chunking and unnecessary AI files such as `llms.txt`. Such a file may serve as voluntary documentation for a particular tool, but it has no established Google advantage and cannot replace accessible HTML, internal links and indexing.

Do not publish hidden model instructions, invented awards, mass pages for prompt paraphrases or synthetic quotations. An EMNLP 2024 paper showed that adversarial text embedded in pages could manipulate some conversational search engines. That is a platform vulnerability, not a durable or acceptable marketing technique.

• No universal structured-data type guarantees a citation; markup must match visible content.
• No supported minimum word, question or statistic count ensures inclusion.
• No cross-platform “AI visibility score” is meaningful without a disclosed protocol.
• Do not refresh a page date without a material factual review.
• Meeting every technical requirement still does not guarantee inclusion.