ChatGPT visibility: how it picks sources, and what you can influence

What ChatGPT visibility means

ChatGPT visibility is how often, and how accurately, ChatGPT names your brand, recommends your product or cites your pages when people ask questions in your category. It comes from two mechanisms: answers from the model’s training data, and answers built from a live web search, which ChatGPT runs when a question needs current information.

The distinction decides almost everything else. A searched answer reflects which pages ChatGPT’s search found and read today, which you can influence with ordinary search work. A trained answer reflects what the model absorbed before it shipped, which you cannot change on any short timescale. Most confusion about ChatGPT visibility, and most bad advice about it, comes from treating the two as one thing.

The audience is large enough to matter for most B2B categories. OpenAI announced in February 2026 that ChatGPT had reached 900 million weekly active users, as TechCrunch reported. Those users do not all research purchases there, but a buyer who asks ChatGPT for a shortlist sees one answer with a handful of names in it, and you are either in it or not.

This guide is about ChatGPT specifically: how it decides to search, how it chooses and cites sources, how to make sure it can read your site, and what to do when it describes your product wrongly. The general principles that apply to every AI assistant are in the pillar guide, what AI search actually retrieves from your page, and the wider work of earning citations across assistants is in how to improve brand visibility in AI search.

How ChatGPT visibility differs from Google rankings

ChatGPT visibility differs from Google rankings in the unit being won, the query being answered and who sees what. Google ranks pages for the query a person typed; ChatGPT, when it searches, runs its own rewritten queries, may personalise them, and presents one answer that names a few brands and cites a few sources.

	Google organic results	ChatGPT answers
What you win	A position on a results page	A mention, a recommendation or a citation inside one answer
Which query counts	The query the person typed	The queries ChatGPT rewrites it into, when it searches
Same for everyone?	Largely, allowing for location and device	No: memory, custom instructions and location can change the search and the answer
Crawler to allow	Googlebot	OAI-SearchBot for search; GPTBot only affects training
Repeatability	Stable enough to rank-track	Brand lists repeat in under 1% of runs of the same prompt (SparkToro study)
First-party reporting	Search Console impressions, clicks and positions	None from OpenAI; referral visits tagged utm_source=chatgpt.com
What you can change	Pages, links, technical health	Searched path: pages and third-party listings. Trained path: only slowly, through consistent descriptions

OpenAI Help Center; Ahrefs, 2025; SparkToro via Search Engine Journal, 2026

Two rows matter most in practice. Because the query that counts is ChatGPT’s rewrite rather than the user’s words, the pages it reads come from a wider cluster than any one keyword, which is why overlap with Google’s top 10 is low. And because there is no first-party reporting from OpenAI, the measurement has to be built by sampling answers and counting referral visits, both covered later in this guide.

What does not change is the base layer. A page ChatGPT cannot crawl, that answers nothing specific, or that nobody else links to or describes, does as badly in ChatGPT search as in Google. Strong conventional SEO is still the starting position for the searched path; ChatGPT adds the access check, the cluster view and the accuracy problem on top.

How ChatGPT decides whether to search the web

ChatGPT decides to search the web automatically when it judges that a question would benefit from current information, and users can also start a search by hand. When it searches, the answer carries citations that link to the sources; when it does not, the answer comes from the model alone.

OpenAI’s ChatGPT search FAQ states both behaviours: ChatGPT “may search the web automatically when your question would benefit from current information”, and a search can be started manually from the tools menu or by typing a slash in the composer. Responses that used search may include citations, and selecting a citation opens its source.

Answered from training

No search runs and no sources are linked. The answer reflects what the model absorbed during training, a snapshot from before it was released. Nothing you publish this quarter changes it.

Answered with search

ChatGPT rewrites the question into its own search queries, reads a handful of results and writes an answer that cites them. This is the path you can influence, mostly with conventional SEO.

A blend of both

ChatGPT searches, then combines what it found with what the model already believed. This is why a page can be cited while the surrounding claims about your product are still out of date.

The practical tell costs nothing: if the answer shows linked sources, a search ran; if it does not, you are reading the training snapshot. Record which happened every time you test, because mixing the two in one set of results makes ChatGPT look random when it is in fact doing two different things.

A simple question about what your brand is can be answered from training, because it does not obviously need current information. Since ChatGPT searches when a question would benefit from current information, asking for current details such as this year’s pricing is the natural way to test the searched path, and starting a search from the tools menu forces it. Testing each path deliberately is more useful than hoping to catch one.

How ChatGPT search picks its sources

ChatGPT search picks sources by rewriting the user’s question into one or more targeted queries, sending them to search providers and reading the results it gets back. Because it searches with its own queries, the pages it cites are often not the ones ranking in Google for the question as typed.

What OpenAI says about the mechanism

OpenAI’s search FAQ gives the outline. ChatGPT sometimes partners with other search providers and “typically rewrites your query into targeted queries” for them; the FAQ links the Microsoft and Shopify privacy policies in that context. The rewritten queries may include a general location estimated from the user’s IP address, though OpenAI says it does not share the IP address itself or account information with providers. Location can therefore change what ChatGPT finds, as it does in any search engine.

Memory changes it too. The Memory FAQ says that when memory is enabled, “ChatGPT may use relevant details to formulate a more useful search query”. Two people asking identical words can send different queries to the search provider and get different sources back.

How little ChatGPT’s sources overlap with Google

In an Ahrefs study of 15,000 long-tail queries published in August 2025, only about 12% of the URLs cited by ChatGPT, Gemini and Copilot ranked in Google’s top 10 for the original prompt, and about 80% of citations did not rank anywhere in Google for it. For ChatGPT’s in-text citations specifically, the figure was 8.0% for Google’s top 10 and 8.1% for Bing’s. Perplexity overlapped most, at 28.6%.

Ahrefs’ explanation is query fan-out plus rank fusion: the assistant searches several variations of the question and merges the results with a method such as Reciprocal Rank Fusion, which rewards pages that appear consistently across variations over a page that ranks first for only one. Ahrefs presents this as the likely mechanism, not as something OpenAI has confirmed.

Data from a tracking vendor gives a sense of scale. Peec AI analysed more than 20 million ChatGPT fan-out queries between October 2025 and January 2026 and found 2.3 to 2.8 fan-out queries per prompt across the five countries it studied, a number that stayed roughly constant, while the average fan-out query roughly doubled in length, from about 6 words to about 12. Peec sells AI visibility tracking, so treat this as one vendor’s data, but the direction is useful: ChatGPT is sending a few, increasingly specific searches per question, not one short keyword.

Worked example (hypothetical)

Reciprocal Rank Fusion, as defined by Cormack, Clarke and Buettcher at SIGIR 2009, scores each page as the sum of 1 divided by (k plus its rank) across the result lists, where k is a constant. The paper fixed k at 60 after a pilot run; take the same value here for illustration, and suppose ChatGPT rewrites a question into three sub-queries, in line with the fan-out counts above.

• Page A ranks first for one sub-query and does not appear for the other two: 1 / (60 + 1) = 0.0164.
• Page B ranks sixth for all three sub-queries: 3 times 1 / (60 + 6) = 3 times 0.01515 = 0.0455.
• Page B scores almost three times higher and is the one read and cited, even though it never ranks first for anything.

The calculator below runs the same arithmetic on your own positions. Enter where two pages rank for each of three sub-queries, with 0 for a page that does not appear. The formula is score = the sum over sub-queries of 1 / (k + rank), skipping any sub-query where the rank is 0. Scores are shown multiplied by 1,000 so the differences are readable.

Constant kPage A rank, sub-query 1Page A rank, sub-query 2Page A rank, sub-query 3Page B rank, sub-query 1Page B rank, sub-query 2Page B rank, sub-query 3

Page A fused score (x 1,000)

16.4

Page B fused score (x 1,000)

45.5

Page B’s score as a multiple of Page A’s

2.8x

Above 1.0x, Page B ranks higher in the fused list

Hypothetical page positions. k = 60 is the value used in the original RRF paper; OpenAI does not publish its own method, so treat the result as an illustration. Enter 0 for a sub-query where the page does not appear.

The operator’s conclusion: ChatGPT visibility on the searched path rewards being a solid result across the cluster of questions around a topic, more than winning one head term. That is the same conclusion the Google AI Overviews data points to, covered in the guide to AI Overviews SEO.

How to make sure ChatGPT can read your site

Making sure ChatGPT can read your site means allowing OAI-SearchBot, OpenAI’s search crawler, in robots.txt and at the firewall. OpenAI says content cannot be included in ChatGPT’s search summaries and snippets if OAI-SearchBot is blocked, and that GPTBot, the training crawler, is controlled separately.

OpenAI’s publisher FAQ also says robots.txt may need updating to give OAI-SearchBot access. It treats training separately: publishers should disallow GPTBot for pages they want excluded from potential training. The FAQ also notes that if OpenAI learns the URL of a disallowed page from a third-party provider or from crawling, it may surface just the link and page title in ChatGPT Atlas, and it states that any public website can appear in ChatGPT search: there is no application or fee for organic inclusion.

The four OpenAI user agents

OpenAI’s crawler documentation lists four user agents and says each setting is independent of the others:

User agent	What OpenAI says it does	Effect of blocking it
OAI-SearchBot	Surfaces websites in ChatGPT’s search features	Your pages stop appearing in ChatGPT search answers, though they can still show as navigational links
GPTBot	Crawls content that may be used to train foundation models	Signals your content should not be used for training; no stated effect on search
ChatGPT-User	Visits pages when a user or a Custom GPT asks ChatGPT to; not automatic crawling	OpenAI says robots.txt rules may not apply, because the visit is user-initiated
OAI-AdsBot	Checks the safety of pages submitted as ads on ChatGPT	Only relevant if you advertise on ChatGPT; its data is not used for training

OpenAI crawler documentation, 2026

Two details from the same page change the access check. OpenAI says a robots.txt change can take about 24 hours to be reflected in search results, so check the next day, not the next minute. And OpenAI publishes the IP addresses each bot uses as JSON files, which is what a firewall or CDN rule should allow, rather than trusting the user-agent string alone.

The access check

1. Open yourdomain.com/robots.txt and look for any group that names OAI-SearchBot. If none does, OAI-SearchBot follows the User-agent: * group, so a broad Disallow there blocks search inclusion.
2. Decide on training separately. Blocking GPTBot keeps pages out of potential training; it is a policy choice and does not, on OpenAI’s description, stop search inclusion.
3. Check your CDN or firewall bot rules. A robots.txt that allows OAI-SearchBot does nothing if a bot-management rule returns a challenge page to it.
4. Search your server logs for the OAI-SearchBot user agent and confirm it receives 200 responses on the pages you care about.
5. Make sure the content that answers questions is in the HTML response, not only rendered by client-side JavaScript after load. Server-rendered text is the safe assumption for any crawler you have not tested.
6. Confirm the pages are indexed in both Bing Webmaster Tools and Google Search Console, for the reasons in the next subsection.

robots.txt set-ups for ChatGPT

Most sites want one of four set-ups. Pick the tab that matches your policy and copy the lines; whether to block GPTBot is a business decision about training, and the OAI-SearchBot lines are the ones that affect visibility. One rule of the Robots Exclusion Protocol, RFC 9309 matters when you paste them: a crawler obeys the group that names it, and falls back to the User-agent: * group only when no group names it. The two are not combined, so any Disallow lines from your * group that should still apply to OpenAI’s crawlers have to be repeated inside their groups. With no group that applies at all, nothing is disallowed.

Search yes, training no

User-agent: OAI-SearchBot
Allow: /

User-agent: GPTBot
Disallow: /

Pages can appear in ChatGPT search answers, and the GPTBot group signals that content should not be used to train foundation models. On OpenAI’s description the two settings are independent, so the training opt-out does not affect search inclusion.

Allow both

User-agent: OAI-SearchBot
Allow: /

User-agent: GPTBot
Allow: /

Pages can appear in ChatGPT search, and content may be used for training. This is also the effect of having no OpenAI-specific groups at all, provided the User-agent: * group, which the crawlers then follow, does not disallow the pages. OpenAI says that when both crawlers are allowed it may use the results of one crawl for both purposes.

Opt out of both

User-agent: OAI-SearchBot
Disallow: /

User-agent: GPTBot
Disallow: /

Pages stop appearing in ChatGPT search answers, although OpenAI says they can still show as navigational links, and content is signalled as not for training. ChatGPT-User visits are made at a user’s request, and OpenAI says robots.txt rules may not apply to them.

Keep one page out

<meta name="robots" content="noindex">

To keep a single page out of ChatGPT search while the rest of the site stays in, OpenAI’s publisher FAQ points to a noindex tag rather than a robots.txt block, and the crawler has to be allowed to fetch the page to read it. A page disallowed in robots.txt cannot show OAI-SearchBot its noindex tag.

Sources: OpenAI crawler documentation and publisher FAQ, linked above. Check CDN and firewall rules as well, since they override robots.txt.

Does ChatGPT depend on Bing or Google?

OpenAI does not say which search providers ChatGPT uses, beyond linking Microsoft’s and Shopify’s privacy policies in its search FAQ. The public evidence is anecdotal and points both ways. In a test reported by Search Engine Journal, Aleyda Solis found ChatGPT could not answer about a new page until it was indexed, and the snippet it used matched Google’s, not Bing’s; the article presents this as one test with no confirmation from OpenAI. Glenn Gabe has been tracking a site that was cited in ChatGPT while invisible in Google and was then deindexed by Bing in August 2026; the case write-up describes the effect on its ChatGPT citations as unsettled.

The practical response does not need the answer. Being properly indexed in both Bing and Google costs one afternoon: verify the site in Bing Webmaster Tools and Google Search Console, submit the sitemap to both, and fix anything either reports as excluded. Bing Webmaster Tools also has an AI Performance report covering Microsoft Copilot citations, described in the guide to tracking brand mentions.

How to improve ChatGPT visibility on the searched path

Improving ChatGPT visibility on the searched path means ranking across the cluster of queries ChatGPT is likely to send, and writing passages it can quote accurately once it reads the page. Most of the work is conventional SEO aimed at the long tail, plus editing for self-contained answers.

1. Map the cluster, not the head term. For each priority question, list the variations ChatGPT is likely to search: the short keyword form, the comparison form, the “for [segment]” form and the follow-up questions. People Also Ask, Search Console long-tail queries and sales-call questions are the best sources.
2. Cover the variations on one strong page or a tight cluster. The rank-fusion example above shows why: steady positions across many variations beat one first place.
3. Answer in the first lines of each section. A heading that matches the question with the answer immediately beneath gives ChatGPT a passage it can use without depending on the rest of the page.
4. Be specific enough to be quoted. A number with its source, a named method, a price, a limit. Generic sentences are absorbed into the summary without attribution.
5. Get described on the pages ChatGPT reads. When a review site, directory or comparison article ranks across your cluster, being listed there accurately is often faster than outranking it.
6. State plain facts about your product on your own site. Pricing model, integrations, who it is for and what it does not do, in sentences that make sense alone.

The Entity Extractor is useful for one narrow check in this list: paste a page ChatGPT cites for your topic and your own equivalent, and compare the specific tools, standards and concepts each names. A page that never names its subject specifically gives ChatGPT nothing precise to carry into an answer.

How to build a list of the third-party pages ChatGPT reads

The fastest gains on the searched path often come from pages you do not own. Collect every URL ChatGPT cites across your test questions, reduce each to its domain, count how often each domain appears in searched answers, and classify the frequent ones (review platform, directory, publisher article, forum thread, competitor, your own site). The full source-gap procedure, with an action for each source type, is in the brand visibility guide linked at the top of this page; the ChatGPT-specific point is to count only the answers in which a search ran, since trained answers cite nothing.

Which sources ChatGPT cites most, and what it rewards

ChatGPT cites community, reference and review sites far more than any brand site: in Ahrefs’ monthly tracking, Reddit and Wikipedia lead its citations. It also leans toward fresher pages than Google’s organic results do. OpenAI documents no role for schema markup or llms.txt.

The most-cited domains

Ahrefs publishes a monthly table of the most-cited websites in ChatGPT, built from its Brand Radar data on a broad sample of US queries. In the September 2026 update, Reddit held 16.8% of citations among the top sources, Wikipedia 7.0%, Consumer Reports 3.7%, Forbes 3.1% and YouTube 2.5%. Ahrefs does not state the number of queries behind the table, and it covers all topics, so a B2B category will look different; it is still the clearest public picture of where ChatGPT’s citations go.

Reddit 16.8%

Wikipedia 7.0%

Consumer Reports 3.7%

Forbes 3.1%

YouTube 2.5%

Ahrefs, most-cited domains in ChatGPT, September 2026 update

The operator’s reading for a B2B brand:

• Community threads are a primary source, not a side channel. If Reddit threads answer your category questions, ChatGPT is likely reading them. Take part where the community’s rules allow it, as a named employee answering the question asked, and never with sockpuppet accounts, which break most communities’ rules.
• Wikipedia follows independent coverage. It is not a page a company can write itself into; if your company is notable enough for an article, accuracy there matters, and if not, the effort belongs in the press and analyst coverage that would make it notable.
• Review and comparison sites carry your category answers. They are where the third-party citation list from the previous section usually points first.
• Your own site still matters for facts. Brand sites are rarely the most-cited domain overall, but they are what ChatGPT reads when the question is about your product specifically.

Freshness

ChatGPT tends to cite newer pages than Google ranks. In an Ahrefs study of about 17 million citations published in July 2025, ChatGPT’s citations were on average 458 days newer by publication date than Google’s organic results, and 58 days newer by last-updated date. Ahrefs’ own caveats matter: cited content was still about 2.9 years old on average across assistants, freshness is one factor among many, and changing a publish date without changing the page is not an update.

The practical rule: when facts on a page change (prices, integrations, product names, data), update the page and its visible date in the same edit. Do not run date-refresh routines on pages whose content has not changed.

Schema markup and llms.txt

OpenAI’s publisher FAQ mentions neither structured data nor llms.txt; its only stated requirements are access for OAI-SearchBot and, to exclude a page, a noindex tag that the crawler is allowed to read. Several ranking guides present schema as a ChatGPT ranking factor without evidence. Keep schema for the reasons it already earns its place in Google, validate it, and do not expect it to move ChatGPT on its own. An llms.txt file costs little to add, but nothing OpenAI publishes says ChatGPT reads it.

ChatGPT shopping, product feeds and ads

ChatGPT has two commercial routes that sit outside organic visibility: product feeds for merchants, and ads. Both are relevant mainly to ecommerce and consumer brands; most B2B SaaS companies can note them and move on.

• Product feeds. OpenAI’s commerce documentation says product feeds give ChatGPT the catalogue data it needs to index products, and that onboarding feeds is currently available to approved partners through an application form, with daily full-feed uploads and API updates during the day. A retailer should apply; a software company has nothing to submit.
• Ads. OpenAI’s crawler documentation describes OAI-AdsBot as checking pages submitted as ads on ChatGPT, which confirms paid placements exist. Ads are bought, not earned, and they do not change the organic answer, so treat them as a paid channel with its own budget and measurement, not as a shortcut to the visibility this guide covers.

Neither route helps with the trained path or with how ChatGPT describes your product in an organic answer. If a shopping or ad listing states a price or capability, keep it consistent with your own site, for the same accuracy reasons covered below.

How ChatGPT’s training data shapes what it says about you

ChatGPT’s training data shapes the answers it gives without searching, and those answers repeat whatever was most widely and consistently said about your brand before the model was trained. The lever is slow and indirect: make the facts about your product consistent and plain everywhere they are published.

If the model absorbed something wrong about your product, it says that wrong thing to everyone who asks a question that does not trigger a search, until a model trained on different data replaces it. There is no correction form for this, and usually no notification that it is happening.

Source	Likely influence	Why
Consistent descriptions across many third-party sites	Strongest	Repetition across independent sources is what a model is most likely to reproduce
Your own documentation stated as plain fact	Strong	Clear, specific and widely crawled
Widely cited coverage or research mentioning you	Moderate	Carries weight if it is repeated elsewhere
Marketing and positioning copy	Weak	Written to be distinctive, so it differs from how everyone else describes the category
Anything published in the last few months	None yet	Not in the training data of the model currently answering

The habit that follows is cheap and slow. Whenever your product is described somewhere you do not control (directory listings, integration marketplaces, comparison pages, review profiles, partner pages), check that the description is accurate and current, and correct it where you can. Individually each fix is trivial; together they are the only influence you have on the trained path.

The one counterintuitive point: a page stating plainly what your product does not do is more likely to be reflected accurately than three pages of benefits, because it contains a fact rather than a claim. Positioning language is the part of your copy least likely to survive summarisation.

How to test ChatGPT visibility without misleading yourself

Testing ChatGPT visibility properly means asking buyer questions in a session with no memory or custom instructions, without naming your brand, several times each, and recording whether a search ran. Testing in your own account is the easiest mistake to make, and the most flattering one.

Trap	What happens	Fix
Testing in your own account	Memory and custom instructions shape the answer and even the search query	Use a Temporary Chat with personalisation off, or search while signed out
Naming your product in the question	It appears because you asked about it	Ask the question a buyer would ask, without your name
Testing once	The same question returns different brands on different runs	Run each question several times and record the share of runs you appear in
Not noting whether a search ran	Trained and searched answers get mixed	Record whether sources were linked for every answer

Each fix is grounded in OpenAI’s own documentation. The Temporary Chat FAQ says unpersonalised temporary chats do not use memory, custom instructions or plugins, and the search FAQ confirms that people who are not signed in can also use web search. On repetition, the SparkToro study reported by Search Engine Journal found ChatGPT returned the same list of brands in fewer than 1% of repeated runs of the same prompt, so a single test is an anecdote.

A ten-minute monthly check

1. Open an unpersonalised Temporary Chat.
2. Ask your three most important category questions, without your brand name, three times each in fresh chats.
3. Ask three brand questions: what your product is and who it is for, how pricing works, and what it does not do.
4. For each answer, note whether a search ran, whether you were named, what was cited, and anything wrong.
5. Compare against the last three months, not last month alone.

That is enough for most companies. When you need a measured trend across several assistants, with run counts sized to detect real changes, the full method is in the guide to tracking brand mentions in AI search.

What to do when ChatGPT gets your product wrong

When ChatGPT gets your product wrong, first check whether the wrong answer came from a search or from training. A searched error is fixed by correcting the page it cited; a trained error is reduced over time by making accurate descriptions consistent across the web.

A wrong answer about pricing, integrations or compliance is delivered confidently, with no competing results beside it to prompt scepticism. If ChatGPT says your product does something it does not, you can get demo requests from people who wanted that capability and a sales team wondering why lead quality fell. That is why the brand questions belong in the monthly check even for companies that do not care about category visibility.

What you see	Likely cause	What to do
Wrong claim, with a linked source	A page ChatGPT searched says it	Open the source. Correct it if it is yours; request a correction if it is a third party’s
Wrong claim, linked source is correct	The model blended old training data with the search	Make the correct fact more prominent and specific on the cited page
Wrong claim, no sources linked	Training data	Fix descriptions across third-party sites; recheck after model updates
Outdated claim (old price, old plan names)	Training data or a stale third-party page	Find stale pages with a site search for the old term and update them

Worked example (hypothetical)

How to measure traffic from ChatGPT

Traffic from ChatGPT can be measured in analytics because ChatGPT adds a utm_source=chatgpt.com parameter to the links it sends, and GA4 now groups recognised AI referrers in an AI Assistant channel. That traffic is a floor, because most answers are read without a click.

OpenAI’s publisher FAQ confirms the tagging: “ChatGPT automatically includes the UTM parameter utm_source=chatgpt.com in referral URLs”. In GA4, Search Engine Journal reported in May 2026 that Google added an AI Assistant channel to the default channel group, naming ChatGPT, Gemini and Claude as examples. The list has not stayed fixed: a July 2026 Search Engine Journal analysis found Google’s live documentation naming ChatGPT, Gemini, DeepSeek, Copilot and Grok, with Claude removed, and Perplexity sessions still landing in Referral, so the channel on its own undercounts assistant traffic.

• Landing pages from ChatGPT. Filter sessions by source chatgpt.com and list landing pages. These are the pages ChatGPT cites and people click, which tells you which content is doing the work.
• Conversion rate of ChatGPT sessions. Compare it with organic search on the same landing pages before deciding how much effort the channel deserves.
• Self-reported attribution. Add “ChatGPT or another AI assistant” to the how-did-you-hear field on sign-up and demo forms. It captures the buyers who read an answer and came back later through search or direct.
• Branded search impressions. A rise in people searching your name can reflect assistant recommendations that never produced a click.

The broader problem of valuing channels that influence buyers without a trackable click, and why last-touch reports undercount them, is set out in measuring organic when attribution is already broken.

A proportionate plan for ChatGPT visibility

A proportionate plan for ChatGPT visibility fixes access and product accuracy first, then strengthens the pages ChatGPT is likely to search for, and only then decides whether ongoing measurement needs a tool. For most B2B companies this is an editing standard and a monthly check, not a separate programme.

Week 1

Access and accuracy

Check robots.txt, firewall rules and logs for OAI-SearchBot, and confirm key pages are indexed in both Bing and Google. Run the ten-minute check. Note every wrong claim about pricing, capabilities and fit, and whether each came from search or training.

Month 1

Fix the sources

Correct the pages behind searched errors. Rewrite your own pricing, integration and limitation pages as plain facts. Start the third-party description audit. Confirm chatgpt.com traffic is visible in GA4 and add the AI option to forms.

Quarter 1

Strengthen the searched path

Map query clusters for your top category questions, fill gaps in coverage, edit key sections so they answer in the opening lines, and pursue accurate listings on the third-party pages ChatGPT cites. Compare three months of checks before deciding on tooling.

Access

robots.txt allows OAI-SearchBot; any User-agent: * group checked for a broad Disallow GPTBot set by a stated training policy, not by accident CDN and firewall bot rules return pages, not challenges, to OAI-SearchBot Server logs show 200 responses to OAI-SearchBot on priority pages Answer text present in the HTML response, not only rendered by client-side JavaScript Site verified and sitemap submitted in both Bing Webmaster Tools and Google Search Console

Testing

Monthly check run in an unpersonalised Temporary Chat Category questions asked without the brand name, three times each Whether a search ran recorded for every answer

Accuracy

Brand questions on pricing, fit and limitations checked monthly Every wrong claim triaged as searched or trained Own pricing, integration and limitation pages written as plain facts Third-party descriptions audited and corrected where you can

Searched path

Query cluster mapped for each priority question Key sections answer the question in their opening lines Accurate listings on the third-party pages ChatGPT cites for your category

Measurement

chatgpt.com sessions visible in GA4, with landing pages reviewed “ChatGPT or another AI assistant” added to how-did-you-hear fields

Ticks are saved in your browser. Use Copy as Markdown to share the gaps.

What is not proportionate: a dedicated ChatGPT programme before you know whether there is a problem, a monitoring subscription before a manual baseline, or rewriting content on the assumption that a specific undisclosed mechanism works a specific way. If you do want the whole job run as one engagement, the AI search visibility service covers access, accuracy and the searched path together.

Accuracy deserves the last word. As assistants become a normal way to research a purchase, what ChatGPT says about your product becomes part of your product surface, like documentation or a status page. Nobody owns it by default and it degrades silently, which is reason enough for the ten-minute monthly check whether or not category visibility ever produces a lead.