Ranking in ChatGPT search without SEO folklore

Ranking in ChatGPT search is not a new version of winning ten blue links. ChatGPT may rewrite one prompt into several searches, gather material from search partners and directly available web content, choose sources that fit each part of the question, and synthesize an answer with some of those sources cited. Your page can be accessible yet never selected, selected as background yet not cited, cited for one prompt yet absent for a close variation. Treating all four states as "ranking" hides the work you need to do.

The practical goal is to make the right page easy to retrieve, easy to verify, and worth citing for a defined class of questions. You cannot buy or guarantee first place. OpenAI says ranking uses several factors intended to find reliable, relevant information and explicitly says top placement cannot be guaranteed. That leaves plenty to improve, but it rules out the usual pitch about a secret ChatGPT ranking factor.

This is the method I would use for a startup or an established B2B company: remove crawler failures, map prompts to answerable claims, publish evidence that another system can quote without guessing, and measure citations separately from visits. Ninety days is enough to establish whether the work is producing a signal. It is not enough to promise permanent visibility in a product that changes its query and sources with context.

Retrieval is a chain, not a position

ChatGPT search builds an answer through several decisions, so there is no universal result position for a domain. The OpenAI Help Center explains that ChatGPT can rewrite a user's request into one or more targeted queries and can send follow-up queries after reviewing initial results. It may use third-party search providers, while OpenAI's launch explanation also says search uses content supplied directly by partners. The original user wording is therefore only the start of retrieval.

Suppose a founder asks, "Which fractional CTO can cut engineering costs without slowing releases?" Search may break that into queries about fractional CTO experience, engineering cost reduction, delivery evidence, location, and perhaps the companies under consideration. A page optimized only for the exact original sentence may lose to separate pages that answer those narrower questions with clear proof.

Keep four states separate when you diagnose performance:

• Discovery means a crawler or provider knows the URL exists.
• Retrieval means the URL enters the candidate set for a particular search.
• Citation means the generated answer visibly attributes a claim to the URL.
• Referral means a person follows that citation or another source link to your site.

A referral implies visibility, but the reverse does not hold. A cited answer may satisfy the reader without a click. A page can also influence synthesis without receiving the visible citation you expected. Do not turn that possibility into a claim that "mentions count" when you cannot observe them. Measure what you can see and label the rest as unknown.

Context also changes results. The Help Center says query rewriting can use general location and relevant ChatGPT Memory when enabled. Follow-up questions carry conversation context. Two people can type the same seven words and receive different sources because their location, memories, prior turns, or product experience differs. A weekly screenshot of one vanity prompt is not a rank tracker.

Search activation is another gate. ChatGPT can choose to search when a request benefits from current web information, and users can also invoke search themselves. A prompt answered without web search has no live result set for you to enter, even if your page would have been an excellent source. Track whether the tested answer actually used search before recording your presence or absence.

The answer is also composed, not copied from a single winning document. One source may support a definition, another a current price, and a third a warning. A page that deserves the price citation may have no reason to appear beside the definition. Evaluate source coverage claim by claim. Asking why a domain did not "win the answer" is usually less useful than asking which claim lacked the strongest available source.

This is why traditional average position does not transfer cleanly. You can still examine which pages and domains recur, but the unit of analysis is a prompt, its search path, an answer claim, and a citation. Any tool that collapses those into one decimal score must disclose the weighting, or the decimal only adds false precision.

Crawl access is the admission ticket

Your content cannot be summarized and cited reliably if OAI-SearchBot cannot fetch it. OpenAI's publisher guidance tells site owners to allow OAI-SearchBot and to make sure the host or content delivery network accepts traffic from OpenAI's published IP ranges. That is a technical prerequisite, not a relevance boost.

Start with an explicit rule instead of assuming a wildcard applies the way you expect:

User-agent: OAI-SearchBot
Allow: /

User-agent: GPTBot
Disallow: /

This configuration allows search discovery while declining potential model training. OpenAI distinguishes OAI-SearchBot from GPTBot, and the publisher FAQ tells publishers to use GPTBot controls for training preferences. Teams routinely blur search retrieval and model training, then block both because legal approved only one sentence in a policy document. Separate the decisions.

After changing robots.txt, test the whole request path. A correct file does not help when the firewall returns 403, bot protection demands JavaScript, the origin rate limits with 429, or a regional rule sends the crawler into a redirect loop. OpenAI's crawler troubleshooting guidance specifically calls out web protection, authentication, CAPTCHA challenges, redirects, and rate limiting as layers worth checking.

Use a real public page and inspect the status, content type, canonical response, and logs:

PAGE='your public article address'
curl -sS -A 'OAI-SearchBot' -o /dev/null \n  -w 'status=%{http_code} type=%{content_type} redirects=%{num_redirects}
' "$PAGE"

A healthy output shape looks like this:

status=200 type=text/html; charset=utf-8 redirects=0

Do not spoof a crawler to bypass a security control you do not own. The command is a diagnostic on your own property. Confirm the published crawler identity and IP ranges in OpenAI's documentation, then correlate the request with CDN and origin logs. Check several page types because a marketing page, documentation route, and JavaScript application can pass through different rules.

Also inspect noindex, canonical tags, authentication, consent walls, and rendered content. OpenAI notes that a disallowed page might still surface as a title and link in some experiences if another provider supplies the URL, but that is not the same as allowing its content in a summary. Partial visibility is a poor operating target.

Prompts reveal the pages you actually need

Build pages around answer units that survive query rewriting, not around hundreds of cosmetic keyword variations. An answer unit is a claim, comparison, definition, procedure, or fact that resolves one part of a user's decision and has enough context to stand alone.

For a fractional CTO service, one giant services page usually tries to answer too much. Searchers ask about cost, engagement length, authority, expected deliverables, conflict with an existing VP of Engineering, security access, and the conditions under which the arrangement fails. Those are different retrieval jobs. A useful page owns one coherent job and links its evidence internally through ordinary site structure, even though this article intentionally contains no links.

Create a prompt inventory from customer calls, sales objections, support threads, on-site search, and language used in recorded demos. Remove prompts that your business cannot answer honestly. Then group the rest by the evidence a strong answer would require. "How much does a fractional CTO cost?" needs ranges, scope boundaries, and pricing assumptions. "Is a fractional CTO safe for due diligence?" needs access controls, responsibility, and an explanation of what remains with the company. One page can cover both only if readers naturally need both at the same decision point.

Do not confuse a prompt with a target phrase. ChatGPT may rewrite "Can two AI engineers replace my development team?" into searches about team size, software delivery throughput, AI coding tools, operating risk, and case evidence. Publish the pieces that can support those searches. Repeating the original question twelve times gives the system less information, not more.

The awkward prompts are often the best editorial backlog. Buyers ask whether the claimed savings depend on layoffs, whether output drops after the transition, whether one engineer becomes a single point of failure, and whether the advisor has done the work in production. A page that answers those questions directly has a clearer retrieval purpose than another optimistic essay about AI productivity.

Query demand still matters. Use conventional search data, customer language, and your own site analytics to decide which questions deserve investment. ChatGPT search does not make ordinary search behavior irrelevant; OpenAI says it works with search providers. You are adding another retrieval surface, not replacing the web beneath it.

Citable pages make claims cheap to verify

The best candidate source lets a reader and a retrieval system verify the relevant claim without reconstructing your argument. Put the direct answer near the start of the section that owns it, state scope and exceptions, name the evidence, and keep the page current. That is good editing before it is any kind of optimization.

Consider two versions of the same assertion. "Our process dramatically improves engineering efficiency" cannot be checked. "We reduced the operating team from 25 people to 2 AI-augmented engineers while maintaining output and uptime" names the baseline, outcome, and constraint. The second still needs supporting context: whose team, which period, how output was judged, what responsibilities moved elsewhere, and who can stand behind the claim. Specificity creates questions, and answering them creates credibility.

Use first-party evidence for facts only you can establish: product behavior, prices, policies, original research methods, and your own operating results. Use independent sources for claims where your commercial interest weakens the evidence. A vendor is the right source for its documented configuration but a poor sole source for declaring its category superior.

Dates matter when a fact changes. Show a meaningful updated date and revise the substance, not just the timestamp. Keep author and organization identity consistent. If a claim depends on a test, describe inputs, method, exclusions, and the output shape. If it depends on a customer result, obtain permission and state what varied. Anonymous superlatives with no method are hard to trust and easy to replace.

OpenAI's own search announcement says ChatGPT search connects users with original, high quality web content and exposes source links. Read that as a product direction, not a disclosed scoring formula. It supports investing in original material and clear attribution. It does not prove that adding an author box, a schema property, or a certain word count raises a hidden score.

Write quotable passages, but do not manufacture slogans. A concise definition should include the boundary people get wrong. A comparison should use the criteria buyers actually apply. A recommendation should state when it fails. Those details make a passage useful in an answer and useful to a human who opens the source.

Negative findings can be strong source material when the method is clear. If a tool failed under a particular repository size, permission model, or deployment constraint, publish the setup and the observed failure. Do not generalize it into "the tool does not work." Precise negative evidence helps a buyer rule out a bad fit, and competing pages often avoid it because honesty feels commercially uncomfortable.

Keep facts and opinions visibly distinct. "The plan costs $5,000" is a factual statement that should match the current offer. "This plan is a better buy than hiring an agency" is a judgment that needs criteria and an identified author. Blending both into anonymous sales copy makes the factual part harder to trust.

Original research needs a denominator, a collection method, and limitations. A chart labeled "teams save 70% with AI" means nothing if readers cannot tell how teams were selected, what counted as savings, and whether failures were excluded. You do not need an academic paper for every operating observation. You do need enough method for another practitioner to decide whether the result applies.

Authority comes from source fit

ChatGPT needs the best source for a particular claim, not the website with the loudest general reputation. A regulator may be the right source for a legal requirement, the vendor manual for a configuration option, the company pricing page for its current fee, and an experienced operator for a judgment call. Source fit explains why one domain can appear for a narrow technical prompt and disappear for a broader market comparison.

This distinction changes content strategy. Do not write a weaker copy of a primary source and expect to outrank it for the fact it owns. Add the part the primary source cannot provide: a tested implementation, a decision framework, a failure analysis, original data with a method, or a comparison based on disclosed criteria. Cite the named source in plain language and explain where its advice stops.

Reputation still affects whether people trust you, but no public OpenAI document gives marketers a portable "domain authority" score for ChatGPT search. Backlinks can help discovery and conventional search visibility, and credible references can expose your work to more people. That does not justify buying irrelevant links or treating every mention as a direct ChatGPT signal.

Entity clarity is similarly practical rather than mystical. Use one consistent company name, describe what the company does without category soup, identify the people responsible for the claims, and keep core facts consistent across pages you control. If third-party profiles contradict your current site, correct the source of the contradiction where possible. Adding ten variations of the company description only creates ambiguity.

Freshness is query dependent. A current price, software release, law, or availability statement needs active maintenance. A durable explanation of a protocol may remain useful for years. Updating every date each month wastes editorial time and teaches nobody anything. Revise when the answer changes, when better evidence appears, or when user language exposes a missing condition.

Structure helps retrieval without replacing substance

Clean page structure makes the answer easier to locate, but markup cannot rescue a weak claim. Use one descriptive page title, a clear opening answer, sentence case headings, short paragraphs, real lists where the count matters, and tables only when rows and columns express a comparison better than prose.

Keep important facts in server-delivered HTML. If the initial response contains an empty application shell and the content appears only after a complex client script runs, crawler access becomes harder to test and less dependable. You do not need to strip the site to plain text. You need the primary content, headings, canonical metadata, and navigation to exist without a fragile interaction sequence.

Structured data should match visible content. Article, Organization, Product, FAQ, and other schema types can remove ambiguity for systems that consume them, but incorrect markup creates conflicting facts. Do not invent reviews, prices, authors, or dates in JSON-LD. Validate syntax, compare it with the rendered page, and remove properties you cannot maintain.

Avoid burying the answer under a biography, a cinematic anecdote, or six paragraphs that define common words. The first sentence after a heading should answer the heading's question. Follow with evidence, conditions, and examples. This structure also makes a page easier for a busy founder to scan, which is a better reason to use it than speculation about passage scoring.

Canonicalization deserves a direct check. If five URLs expose nearly identical text through tracking parameters, print views, category paths, and localization mistakes, you force retrieval systems to choose among duplicates. Pick the intended URL, redirect true duplicates when appropriate, and keep canonical tags and internal references consistent. Do not canonicalize genuinely localized pages to English; each language page should be a complete, useful version with correct language signals.

Page experience matters at the failure extremes. A mobile overlay that blocks the answer, a page that takes repeated retries, or a consent loop can prevent access and repel readers. I would fix those defects. I would not tell a founder that changing a button color is a ChatGPT ranking tactic.

Measurement must separate citations from clicks

Track observable outcomes by prompt class, page, and date rather than reducing the project to one rank number. OpenAI's publisher FAQ says referral links from ChatGPT include utm_source=chatgpt.com, which gives you a clean starting filter in analytics. Keep referral records at the landing-page level and preserve the query context you can collect lawfully. OpenAI does not send you the user's private conversation.

Citation monitoring needs a controlled test set. Create prompts for discovery, comparison, objection, and purchase intent. Run them in fresh conversations on a consistent schedule, record whether search activated, capture the cited sources and answer claims, and note material context such as location. Repeat a subset with realistic follow-up questions. The result is an observation log, not a promise that every user saw the same answer.

Your weekly record can stay small:

• Prompt and intent class
• Test date and relevant context
• Cited domains and exact pages
• Your page's citation status and supported claim
• Referral sessions and qualified actions for the landing page

Do not score an unlinked brand mention as a citation. Do not call a referral a lead unless it reaches your lead definition. Do not celebrate more ChatGPT sessions if they bounce from a page whose claim does not match the answer. The useful chain is citation, qualified visit, meaningful action, and commercial outcome, with honest gaps where identity cannot be connected.

Compare changes in batches. If you open crawler access, rewrite every page, change analytics, and launch a public relations campaign in the same week, you cannot tell what moved. Technical access comes first because everything else depends on it. Content experiments should then change one page group or answer pattern at a time.

Connect measurement to the page's actual job. An informational definition may earn citations and assist later sales without producing many direct bookings. A pricing comparison near purchase intent should produce more qualified actions if it matches the visitor's question. Use different success thresholds for those pages instead of forcing every citation into the same conversion target.

Preserve evidence when results change. Save the response text, cited page, test context, and page version or publication date. A screenshot alone can prove appearance, but it cannot tell you which version of your claim was live. This record becomes useful when a citation vanishes after an edit or when an old answer keeps circulating after the source changed.

Expect volatility. Search providers update indexes, ChatGPT rewrites prompts, sources publish new evidence, and the product changes. A drop on one prompt can be noise. A repeated loss across a prompt class, paired with a competing source that answers the question better, is editorial information.

A 90 day plan should produce evidence

A useful 90 day program leaves you with crawl proof, a focused source library, and a measurement record. It should not end with a slide claiming "AI visibility improved" because somebody saw the brand once.

1. Days 1 to 30: establish access and a baseline. Inventory indexable pages, check OAI-SearchBot rules, test CDN and origin responses, inspect noindex and canonicals, and record current ChatGPT citations for 30 to 50 prompts. Group those prompts by intent and the evidence required. Choose five to ten page opportunities where the company can add a primary fact, tested procedure, or experienced judgment that already exists inside the business.

2. Days 31 to 60: publish source-grade answers. Repair the access defects first. Update or create the selected pages with direct opening answers, named authors, scope, dates, proof, and explicit limits. Add structured data only where it reflects visible facts. Ask sales, support, engineering, and leadership to challenge each page: if they can produce an obvious exception or unsupported claim, fix it before publication.

3. Days 61 to 90: observe, compare, and expand carefully. Re-run the prompt set weekly, review cited competitors, and inspect ChatGPT referrals and downstream actions. Improve pages where another source wins by being clearer, more current, or better evidenced. Expand only the answer patterns that show qualified citations or useful referrals. Keep a control group of unchanged pages so you do not attribute every movement to your work.

Assign owners before day one. Engineering owns crawler access and logs. Editorial owns answer quality and update dates. A subject expert owns factual sign-off. Marketing operations owns the prompt log and referral reporting. One person should reconcile the results each week, because four separate dashboards can hide a broken chain.

Set decision rules while enthusiasm is low. For example, continue a page pattern when it earns repeat citations across several related prompts or produces qualified referral behavior. Revise it when the cited claim is wrong, stale, or mismatched. Retire it when the business has no credible evidence for the question. These rules prevent the team from preserving weak pages because someone spent a month writing them.

Budget for maintenance. Pricing, product capabilities, leadership details, and time-sensitive comparisons need owners and review intervals. If the company cannot maintain a claim, narrow it or remove it. Stale precision is worse than an honest boundary.

Stop buying certainty that does not exist

No consultant can guarantee that your page will rank first in every ChatGPT search result. OpenAI says this directly, and the retrieval process makes the promise incoherent: prompts are rewritten, context varies, and sources compete at the claim level. A guaranteed placement pitch usually repackages ordinary SEO work, automated mention checks, or mass-produced question pages under a new acronym.

Programmatic pages are tempting because prompt variations look infinite. Most companies should resist. If 500 pages repeat the same thin answer with a different industry or city inserted, they do not add 500 pieces of evidence. They create a maintenance burden, duplicate claims, and plenty of chances to contradict the core site.

Do not pay for fabricated authority. Irrelevant guest posts, synthetic reviews, fake expert profiles, and unsupported comparison tables may create discoverability for a while, but they make verification harder and expose the company to reputational and legal risk. Search visibility that depends on nobody checking the source is not an asset.

Also reject reports that hide the prompt set. A vendor can make a visibility score rise by choosing favorable prompts, adding your brand to the question, or counting mentions that were not citations. Require the exact prompts, context, dates, cited pages, and scoring rules. You should be able to reproduce a sample manually.

At oleg.is, I would put this work inside a Team & AI Audit only when search visibility connects to engineering capacity, content operations, and measurable savings; otherwise it becomes another isolated marketing project. The first action is simpler: fetch your ten most important source pages as OAI-SearchBot, then ask whether each page contains one claim a skeptical buyer could verify. If either test fails, you have found work worth doing before anyone invents a ranking score.