SEO for AI search starts with better evidence

When a buyer asks ChatGPT for a shortlist, your page is no longer competing only to rank for the sentence they typed. It is competing to supply a defensible piece of the answer: a price condition, a comparison, an implementation risk, a definition, or evidence for a recommendation. That changes the unit of SEO work from one keyword and one page to a set of claims that an answer engine can retrieve, verify, and cite.

Traditional search still matters. ChatGPT Search, Google AI features, and Microsoft Copilot all depend in different ways on search indexes, crawling, and accessible pages. You do not get to skip technical SEO because the result now reads like prose. But a technically sound page full of generic advice gives an answer engine little reason to select it over a primary source, a sharper specialist, or a page with concrete proof.

SEO for AI search is therefore an editorial and evidence problem built on top of normal SEO. The practical job is to map the questions behind a buying decision, publish the smallest complete set of pages that answers them, and make every consequential claim easy to extract without stripping away its conditions.

The unit of competition is a claim, not a keyword

AI answers assemble decisions from smaller retrieved claims, so a page can influence a buyer without becoming the main result for the buyer's original wording. A founder might ask for the best incident management platform for a small regulated company. The answer engine may search separately for audit logs, data residency, on-call pricing, implementation time, integrations, and recent product changes. A vendor can be absent from one broad result and still become the cited source for a decisive constraint.

This is the sharp distinction many content teams blur: topic coverage is not claim coverage. A 4,000-word guide can mention every subtopic while proving none of its statements. Conversely, a focused page can own one important claim because it states the answer, defines the boundary, gives the evidence, and records when the evidence was checked.

Classic keyword research usually starts with phrases and volumes. Keep that data, but add a decision model. For every commercial topic, identify the buyer, the job they need done, the constraints that could eliminate an option, the comparisons they will make, and the proof they would accept. Those dimensions produce the retrieval questions hiding inside a conversational prompt.

Use a claim record during planning:

This record prevents a common failure: the marketing page says the product is easy, the documentation describes twelve setup stages, and a review page supplies the only specific staffing estimate. An answer engine has good reasons to cite the review instead of the vendor. The problem is not a missing AI schema. The vendor never published the useful claim with its limits.

Query fan-out turns one prompt into a retrieval map

Query fan-out means an answer system can issue several related searches across subtopics and sources before composing its response. Google Search Central uses that exact term for AI Overviews and AI Mode. OpenAI's ChatGPT Search documentation gives the same behavior in plainer terms: it may rewrite a question into one or more targeted searches, inspect the results, and then issue more specific searches.

Take this buyer prompt: Which customer support platform should a 40-person SaaS company choose if it needs EU data handling, a quick migration, and predictable costs?

A useful fan-out map might contain these branches:

1. customer support platforms for a 40-person SaaS company
2. EU data location and subprocessors for each candidate
3. migration path from the current platform
4. seat, usage, and add-on pricing conditions
5. administrator workload after rollout

The list is not a set of phrases to repeat. It is a set of research tasks. Each branch can require a different source type. The vendor's security page may support the data claim, product documentation may support migration steps, a pricing page may support cost, and an experienced operator may be the best source for administrative workload.

Build maps from real sales material, not imagination alone. Pull questions from discovery calls, objections in CRM notes, support tickets before purchase, request-for-proposal documents, comparison searches, and the follow-up prompts people use in AI tools. Remove questions that do not change a decision. Add awkward questions that sales copy avoids, such as what fails during migration, which cost grows fastest, or when the product is a poor fit.

One page does not need to rank for every branch. Your site does need a coherent answer for each branch you have authority to address. Connect the pages with descriptive internal links and consistent terms so a crawler can discover the relationship. If two pages answer the same branch with slightly different facts, consolidate them or assign each a clear audience and purpose. Near duplicates make the preferred source harder to identify.

Do not create a page for every generated question. That popular recommendation turns a useful map into thin programmatic content. Group questions when the same evidence answers them. Split a page when the buyer intent, evidence, owner, or update cycle differs. Page count follows information ownership, not the number of rows in a keyword export.

Citations go to evidence that survives extraction

A citation-worthy passage states a claim together with enough context for the claim to remain true when quoted or summarized. Answer engines do not owe your homepage a citation because your company has experience. They select sources that support the sentence they are generating.

The strongest evidence is usually close to the fact: official documentation for behavior, a maintained pricing page for commercial terms, a standards document for requirements, a reproducible test for performance, and a named practitioner for an operating judgment. Third-party analysis still matters, especially for comparisons and failures, but it should show its method.

I use a four-part evidence test:

• Can a reader identify the exact claim without reading three preceding paragraphs?
• Does the page show how the author knows it?
• Are conditions, exclusions, and units next to the claim?
• Is there an owner and a visible review date for facts that can change?

Consider two versions of the same statement. Our migration is fast and easy contains no usable fact. A standard migration covers ticket history, users, and macros; custom app data needs a separate export gives an answer engine a bounded claim. Add the actual procedure, sample output, and known failure conditions, and a buyer can verify it.

Named sources need engagement, not decoration. Google Search Central says there are no extra technical requirements for inclusion in AI Overviews or AI Mode beyond normal Search eligibility. I agree with the technical point and reject the lazy editorial inference. No special tag is required, but selection still depends on whether a page answers a retrieved subquestion better than competing sources. Ordinary eligibility opens the door; specific evidence gives the system a reason to walk through it.

OpenAI's publisher guidance says sites should allow OAI-SearchBot if they want content included in ChatGPT summaries and snippets. That is actionable, but it does not promise placement. OpenAI explicitly says ranking depends on several relevance and reliability factors and offers no guaranteed top position. Treat crawl access as a prerequisite, not a growth strategy.

Avoid invented precision. If your company has not measured setup time across a defined sample, do not publish an average because competitors use one. Describe the known stages and variables instead. A sober boundary often earns more trust than a suspiciously neat number.

Pages must answer cleanly without becoming answer fragments

The content shape that survives AI retrieval is a self-contained answer followed by proof, nuance, and a path to action. This is good writing for humans as well. A buyer scanning a page wants the conclusion before the history lesson, while a serious evaluator needs the assumptions behind it.

Start each major section with the answer it owns. Use descriptive headings that match genuine subquestions. Put definitions near the terms, label comparison dimensions consistently, and keep important caveats in the same paragraph or table cell as the claim. If a caveat lives in a footnote or another page, extraction can detach it.

Four shapes repeatedly work because they match decision tasks:

• A comparison table with explicit criteria and a short explanation of tradeoffs
• A procedure with prerequisites, expected output, and recovery from failure
• A definition that separates terms people routinely confuse
• A decision rule that says who should choose what and under which conditions

Tables help only when the cells contain actual information. A grid of checkmarks hides degree, limits, and evidence. Write EU region available for application data; support metadata follows separate terms rather than marking a generic compliance row. The longer cell is more honest and more extractable.

FAQs can capture late-stage questions, but they should not carry facts that deserve full treatment. A two-sentence answer about data retention cannot replace a maintained retention policy. Use FAQs for concise clarification and route the evidence to the page that owns it.

Schema markup is similar. Valid structured data can clarify entities and page types, but it cannot rescue vague copy. Do not add unsupported review, author, or FAQ markup in the hope of manufacturing authority. The visible page, the structured representation, and the underlying evidence should agree.

The worst response to AI search is to compress every page into short answer blocks. That removes the experience, method, examples, and exceptions that distinguish the source. Put the direct answer first, then earn it. Retrieval needs a clean passage; buyers need enough depth to decide whether it applies.

Technical eligibility still decides whether evidence can be found

AI visibility starts with crawlable, indexable, renderable pages and a clear canonical version. Google says a page must be indexed and eligible to show a snippet before it can appear as a supporting link in its AI features. OpenAI tells publishers to allow OAI-SearchBot for inclusion in ChatGPT search summaries and snippets. Neither statement supports abandoning sitemaps, internal links, canonicals, status codes, or server-side access checks.

Audit crawler policy deliberately. Search retrieval and model training are different uses, and crawler names reflect that distinction. A team may choose to allow search discovery while restricting training. Do not copy a blanket block from a social post and assume every AI-related user agent does the same job.

A minimal policy for ChatGPT search discovery looks like this:

User-agent: OAI-SearchBot
Allow: /

Sitemap: https://example.com/sitemap.xml

This fragment does not grant ranking, force a citation, or describe other crawlers. Check OpenAI's current crawler documentation and published IP ranges, then confirm that your CDN, web application firewall, and origin return the intended response. A permissive robots file does nothing if a bot receives a challenge page or a 403 response upstream.

Run a small release check for every important evidence page:

URL: canonical HTTPS URL
HTTP: 200 without login or bot challenge
Indexing: index allowed
Snippet: snippet allowed at the intended length
Canonical: self or deliberate parent
Rendered claim: present in returned HTML
Updated: visible and accurate

Keep important facts in HTML rather than only inside images, video, or client-side widgets that fail under restricted rendering. Make internal links real anchors with descriptive text. Include changed pages in the sitemap and notify supported engines through their documented mechanisms. Then inspect server logs. A dashboard can say a page is eligible while the origin shows repeated failures.

Preview controls deserve special care. Google documents nosnippet, data-nosnippet, max-snippet, and noindex as controls for how page content can appear in Search, including its AI features. Bing also supports controls for limiting material used in search and AI answers. Restrict sensitive or paid material deliberately, but recognize the tradeoff: content an engine cannot show may be less useful as support for an answer.

Migrate the site by decision cluster, not by traffic rank

An existing site should migrate in clusters tied to buyer decisions, because updating isolated high-traffic pages leaves contradictions around them. Traffic rank tells you where visits occurred. It does not tell you which pages influence a shortlist, resolve an objection, or supply a cited fact.

Start with one commercial decision that matters to revenue. Inventory every page that claims to help with it, plus the documentation, policies, pricing, and case material that contain supporting facts. For each URL, record its intended audience, claim owner, evidence type, last factual review, canonical status, organic traffic, assisted conversions, and known AI referrals or citations.

Score each page on five questions, using zero for absent, one for partial, and two for complete:

Do not sum the score blindly. A page with perfect structure and false pricing still fails. The rubric exposes where editorial polish is hiding a factual gap.

Assign one of four actions: keep, repair, merge, or retire. Keep pages that own a distinct question and remain correct. Repair useful pages with weak evidence or shape. Merge pages that compete for the same intent and rely on the same facts. Retire pages that answer no current decision, while redirecting only when a true replacement exists.

Work outward from the source of truth. Fix pricing, policy, documentation, and product facts before rewriting commentary about them. Then update comparison and educational pages. If five articles repeat a fact that changes monthly, replace the copies with a concise explanation and point readers toward the maintained source, using plain internal link text.

Preserve URLs that already have links, qualified traffic, or a clear role unless the information architecture is genuinely wrong. AI search is not a reason to rename every slug or move the blog. Large migrations create crawl noise and destroy useful history. Change the content model first; change URLs only when consolidation demands it.

A twelve-week rollout should leave evidence, not a slide deck

A practical rollout can cover one decision cluster in twelve weeks, produce measurable changes, and teach the team how to repeat the work. The calendar is not a promise of ranking. It is an operating constraint that stops the audit from expanding across the whole site.

During weeks one and two, collect buyer questions and build the fan-out map. Interview sales, support, product, and two people who recently evaluated the category if you can reach them. Mark which questions change a decision and which team owns the answer. Capture the current answer, not the answer marketing wishes were true.

During weeks three and four, inventory and score the cluster. Crawl the relevant paths, review analytics and search data, test crawler access, and find conflicting claims. Select a small set of source pages that must become authoritative. Every selected page gets an accountable owner.

During weeks five through eight, repair evidence and rewrite. Create missing artifacts, such as a migration output sample, an architecture diagram described in text, a pricing example with assumptions, or a failure recovery procedure. Update source pages first. Rewrite supporting pages around distinct questions and merge overlaps.

During weeks nine and ten, release and verify. Check canonical tags, rendered HTML, robots rules, sitemaps, redirects, snippets, and analytics attribution. Ask people outside the project to answer the mapped buyer questions using only the published pages. Any answer that requires oral context is still missing from the site.

During weeks eleven and twelve, observe retrieval and conversion behavior. Repeat a fixed set of prompts across the AI search products relevant to your market, saving the answer, citations, date, locale, and account conditions. Review referrals and sales conversations. Decide the next cluster based on business value and evidence gaps, not whichever prompt produced the most flattering screenshot.

The deliverables are concrete: a versioned question map, a claim register, repaired source pages, redirect records, access test results, and a baseline observation set. If an agency hands back only recommendations, the company still owns the same broken content system.

Measure citations as observations and revenue as the outcome

AI search measurement is incomplete, so separate what you directly observe from what you infer. A citation proves that a system displayed your source for one answer under recorded conditions. It does not prove stable ranking, agreement with your claim, or influence on a purchase.

Maintain an observation log with these fields:

{"prompt_id":"support-platform-eu-01","surface":"named product","locale":"en-US","observed_at":"ISO-8601 timestamp","cited_url":"canonical URL or null","claim_supported":"short label","answer_saved":true}

Run the same prompt set on a schedule and after material page changes. Keep the prompts stable enough to compare, but add new prompts when sales evidence reveals a new decision branch. Record personalization and location when they may affect results. Never report a citation share without stating the prompt set and method.

Microsoft's Bing Webmaster Tools now exposes AI citation activity, cited pages, and samples of grounding query phrases for supported Microsoft experiences. Its documentation carefully says citation counts do not indicate placement, importance, or the role a page played in an individual answer. That caveat should shape every internal dashboard, including data from other providers.

Use three measurement layers:

• Technical: crawler access, indexability, canonical selection, and successful fetches
• Retrieval: observed citations, cited URLs, supported claims, and query branches
• Business: qualified referrals, assisted opportunities, sales mentions, and conversions

Referral traffic will undercount influence because many buyers consume the answer without clicking, return later through another channel, or mention the brand in a call. Do not solve that gap by inventing an attribution multiplier. Add a plain discovery question to forms and sales calls, preserve self-reported answers, and compare the language buyers use with your fan-out map.

Search Console includes Google's AI feature traffic inside the general Web search reporting rather than offering a clean AI-only view. That limits diagnosis. Pair page-level search changes with your controlled observations and conversion data, and label conclusions as inference when the source does not identify the AI surface.

The useful executive report is short: which decision clusters gained or lost cited coverage, which source pages supported those citations, what qualified behavior followed, and which factual gaps block the next improvement. A giant prompt leaderboard invites teams to optimize for screenshots.

Commercial pages need boundaries buyers can trust

Buyers use AI search to compress vendor research, so commercial pages must make disqualifying facts as clear as benefits. Hiding limits may preserve a form fill, but it weakens the source and wastes sales time with poor-fit prospects.

Comparison pages should define the compared buyer, use consistent dimensions, separate verified facts from judgment, and say when each option wins. A page titled as one product versus another that never recommends the competitor is advertising, not analysis. An answer engine can retrieve better-balanced sources.

Pricing content needs units, assumptions, required add-ons, and change ownership. If a quote depends on usage, show a worked example without pretending it applies to everyone. If pricing requires a call, explain what inputs shape the quote. Buyers ask AI tools to estimate total cost precisely because vendor pages avoid this work.

Case studies should expose the mechanism between action and result. Name the starting condition, intervention, elapsed period, measurement method, and factors the team did not control. Do not turn one customer's result into a universal promise. If confidentiality prevents useful detail, publish a technical pattern or anonymized process without invented color.

State poor-fit conditions. A small team may not need the governance model built for a large enterprise. A company with no owner for a complex platform should fix ownership before buying it. These sentences may reduce raw leads and improve qualified conversations.

For founders who lack an internal owner for this migration, a Team & AI Audit can identify where the current team and AI workflow waste engineering capacity before the company commits to a broader transformation. Keep the content decision cluster in scope only if it connects to engineering work; the audit is not a substitute for an SEO specialist or customer research.

Publishing ownership determines whether citations last

Durable visibility comes from maintaining claims when reality changes. Most sites do not fail because writers cannot format a table. They fail because nobody owns the fact in the table after launch.

Attach each volatile claim to a source and a trigger. Pricing changes should trigger pricing examples and comparison reviews. A product deprecation should trigger documentation, migration guidance, and affected commercial claims. A policy change should trigger every page that summarizes the policy. Time-based review dates help, but event-based triggers catch changes earlier.

Use a lightweight claim register rather than another sprawling content calendar. Each record needs the claim, source page, evidence owner, pages that reuse it, volatility, last check, and next trigger. Automation can flag dependencies, but a named person must decide whether the statement remains true.

Editors should reject drafts that merely remix public sources. A publishable page needs at least one thing the company can uniquely know or demonstrate: first-hand operating judgment, a reproducible artifact, a maintained primary fact, a transparent comparison method, or a failure explained with enough detail to prevent repetition. Generating more prose does not fill an evidence gap.

Keep classic SEO hygiene in the release process. Titles, metadata, internal links, canonicals, structured data, and accessibility still matter. The change is that editorial review now asks another question: if an answer engine extracted this passage to support a buyer's decision, would the statement remain accurate and properly bounded?

Do not wait for perfect attribution. Choose one high-revenue decision cluster, publish the evidence your sales team already explains in private, and record the retrieval baseline. In twelve weeks you should own a better source of truth even if citation behavior barely moves. If the site still depends on a writer asking product managers the same questions next quarter, the migration did not happen.