Ask an assistant “what’s the best invoicing tool for freelancers” and the answer names three to five brands. Those brands get a head start that most marketers can’t see in their analytics. The assistant picks them from two places: what the model absorbed in training, and what it finds by searching at the moment of the question. You have far more say over the second, mostly through third-party pages, reviews and clear facts on your own site. For the step-by-step work of improving visibility, see the AI search visibility tactic.
Two layers: what the model knows and what it fetches
The first layer is the model’s own knowledge. During training a model reads an enormous amount of public text, and if your brand was described widely and consistently before the training cut-off, the model “knows” it. Launched last year? It may not. You can’t edit this layer, and changes only appear when a new model is trained.
The second layer is live retrieval. For many questions, especially about products, prices or anything recent, the assistant runs searches, reads a handful of pages and writes its answer from them. This layer changes daily, and it is where most of your influence lies.
OpenAI’s shopping documentation spells it out. When choosing which products to show, ChatGPT considers structured metadata from first-party and third-party providers, such as price and product description, other third-party content, and “model responses generated by ChatGPT before it considers any new search results”. Put plainly: prior knowledge plus fresh retrieval, filtered by safety and product policies.
How each assistant finds its sources
The four big assistants differ mainly in where their live retrieval comes from. This is what the official documentation says as of October 2026.
| Assistant | Where live answers come from | What controls your eligibility | What you can measure |
|---|---|---|---|
| ChatGPT | Rewritten queries sent to search partners, plus OpenAI’s own search crawler | Allow OAI-SearchBot and OpenAI’s published IP ranges | Referral visits (GA4 AI Assistant channel), your own prompt audits |
| Gemini | Google Search index, supplied to the model at prompt time (“grounding”) | Normal Google indexing; Google-Extended is a separate training and grounding control | Referral visits, prompt audits |
| Perplexity | Its own index built by PerplexityBot, plus user-triggered page fetches | Allow PerplexityBot and Perplexity’s published IP ranges | Referral visits, prompt audits |
| Microsoft Copilot | Bing’s index | Normal Bing indexing | Bing Webmaster Tools AI Performance report (citations), referral visits |
The eligibility column is the cheapest thing on this page to fix. If your robots rules or bot protection block OAI-SearchBot or PerplexityBot, nothing else you do for those two assistants will count, so check them before anything else.
ChatGPT
OpenAI’s help article on web search says ChatGPT “sometimes partners with other search providers” and typically rewrites a question into one or more targeted queries for them. The page links to Microsoft’s and Shopify’s privacy policies as examples of providers. It also says ChatGPT “ranks search results using multiple factors intended to help users find relevant, reliable information”, and that placement is not guaranteed.
To be eligible, OpenAI says, you must let OAI-SearchBot crawl your site and make sure your host or CDN allows its published IP addresses. Its crawler documentation adds that GPTBot, which relates to model training, is a separate setting. Blocking GPTBot does not remove you from search answers. Blocking OAI-SearchBot does.
On products, the same shopping article says merchants are ranked on factors like availability, price, quality and whether they are the maker or primary seller. Shopify merchants are already integrated through Shopify Catalog; other merchants can apply to send a direct product feed.
Gemini
Google’s crawler documentation describes grounding as “providing content from the Google Search index to the model at prompt time”. In effect, the pages Gemini can draw on are the pages Google has indexed. The Google-Extended token in robots.txt controls whether your content is used for Gemini training and grounding, and Google says it “does not impact a site’s inclusion in Google Search nor is it used as a ranking signal in Google Search”.
Perplexity
Perplexity’s documentation says PerplexityBot “is designed to surface and link websites in search results on Perplexity” and is not used to crawl content for training foundation models. A second agent, Perplexity-User, visits pages when a user’s question needs them. Because a person triggered the visit, it generally ignores robots.txt rules.
Copilot
Microsoft launched its AI Performance report as a public preview in Bing Webmaster Tools in February 2026. It shows how often your pages are cited in Microsoft Copilot, AI-generated summaries in Bing and select partner integrations: total citations, the average number of your pages cited, and a sample of “grounding queries”, the phrases the AI used when it retrieved your content. There are no click figures. Even so, it is a rare first-party view of how often an assistant cites you.
What the evidence says about which signals count
No assistant publishes a ranking formula. What we have is official guidance and a few correlation studies.
Ahrefs studied 75,000 brands to see what correlated with being mentioned in Google’s AI Overviews. In its May 2025 analysis, branded web mentions had the strongest correlation (0.664), followed by branded anchors (0.527) and branded search volume (0.392). Domain Rating scored 0.326 and number of backlinks 0.218. Ahrefs stresses that this is correlation, not causation, and that it only covered domains above a certain authority level. It still fits how retrieval works. An assistant that reads five pages about “best invoicing tools” can only name brands those pages mention, so third-party mentions outweigh links.
Reviews and forums feed product answers. OpenAI says its product review summaries are based on reviews from public websites, and that labels such as “Budget-friendly” may reflect what reviewers frequently mention. Thin or old reviews leave the assistant little to summarise, which is why review site presence and real forum answers are worth the effort.
Clear, consistent facts cut down mistakes. In its AI Performance announcement, Microsoft recommends clear structure with headings, tables and FAQ sections, evidence behind claims, current content, and less ambiguity, so that text, images and video describe the same entities consistently.
Structured data helps most where the facts are structured. For shopping, OpenAI explicitly names structured metadata such as price and description, and product feeds. For general brand questions, Google says no special markup is needed for its AI features. Use schema to state facts clearly. It won’t switch on mentions by itself.
What you control, ranked by influence
- What third parties say about you: reviews, comparison articles, “best of” lists, forums, industry publications.
- How clearly your own site says what you are and who you serve, on the homepage, pricing and use-case pages.
- Whether crawlers can reach you at all (robots rules, bot protection, server-rendered content).
- Product data quality if you sell physical products, meaning feeds, prices and availability.
- The model’s training data. Real, but slow and indirect, and it follows from the first two over time.
Why tracking AI mentions is harder than tracking rankings
You run “best invoicing tool for freelancers” in ChatGPT, see your brand in fourth place, and report it. A week later you run it again and you’re missing. Did something change? Probably not.
SparkToro, working with an AI-tracking company, asked 600 volunteers to run 12 prompts through ChatGPT, Claude and Google’s AI a combined 2,961 times in late 2025. Its January 2026 write-up found a less than 1 in 100 chance that ChatGPT or Google’s AI would return the same list of brands across two runs. Lists in the same order turned up roughly once per 1,000 runs. Yet some brands were stable: across 994 responses to human-written headphone prompts, the four leading brands appeared in 55% to 77% of answers. The authors note the work was not peer reviewed and disclose the commercial partnership.
So a brand’s position in a single answer tells you almost nothing. How often it appears across many answers does.
How to track brand mentions in AI answers
Start with a fixed prompt set of 30 to 100 prompts, grouped by type: category (“best X for Y”), comparison (“X vs Y”), problem (“how do I stop late payments”) and brand (“is X any good”). Write them the way buyers talk, using wording lifted from sales calls and support tickets.
Run each prompt several times per assistant per cycle, logged out or in a clean profile, from the countries you sell in. A paid tracking tool can automate this; for a small set, a spreadsheet is enough.
Then log four numbers:
- mention rate, the share of runs where your brand appears
- share of voice, your mentions divided by all brand mentions in your category prompts
- description accuracy, the share of mentions that get your pricing, audience and features right
- cited sources, the URLs the assistant links to when it mentions you and when it mentions competitors
Put first-party data alongside it. GA4’s AI Assistant channel shows referral visits, Bing’s AI Performance report shows Copilot citations, and a “How did you hear about us?” field on sign-up and demo forms catches the rest. Plenty of people read an AI answer and search for you later, and no referral report captures that.
Review monthly, not daily. Compare month-on-month mention rate across the whole prompt set and ignore movement on single prompts. If buyers in your category compare options before they buy, and most do, this audit belongs next to your search and paid reports. It takes a few hours a month and shows you where buyers are being steered before they reach your site.
Turn the audit into a to-do list
The cited-sources list is the most useful thing the audit produces. Picture a competitor mentioned in 60% of runs and you in 15%, with the assistants citing the same two review sites and one comparison article every time. That gives you three jobs: get properly listed on those review sites, make sure the comparison article has accurate facts about you, and publish a fair comparison page of your own. If assistants describe your pricing wrongly, find the page they cite and fix it there.
Questions people ask
Can I pay to be recommended in ChatGPT answers?
Not in the organic answer. OpenAI's help centre says product results in ChatGPT are selected independently, are not ads and are not influenced by OpenAI partnerships. Ads in ChatGPT exist but are shown separately from product results.
Does blocking GPTBot remove my site from ChatGPT search?
No. OpenAI's crawler documentation says GPTBot relates to model training and OAI-SearchBot relates to search, and that each setting is independent. Sites that block OAI-SearchBot are not shown in ChatGPT search answers, apart from plain navigational links.
Why does ChatGPT give a different list of brands every time?
Answers are generated, not looked up. A January 2026 SparkToro study of 2,961 runs found a less than 1 in 100 chance that ChatGPT or Google's AI would return the same brand list twice for the same prompt. That is why visibility should be measured across many prompts and repeated runs.
What is the best metric for brand visibility in AI answers?
The share of answers that mention your brand across a fixed set of prompts, each run several times. Ranking position inside an answer changes too much between runs to be useful on its own.
Sources
- OpenAI Help Center: Searching the web with ChatGPT
- OpenAI Help Center: Shopping with ChatGPT Search
- OpenAI: Overview of OpenAI crawlers
- Perplexity: Perplexity crawlers
- Google Search Central: Google's common crawlers (Google-Extended)
- Bing Webmaster Blog: Introducing AI Performance in Bing Webmaster Tools public preview
- Google Analytics Help: Default channel group
- Ahrefs: An analysis of AI Overview brand visibility factors (75K brands studied)
- SparkToro: AIs are highly inconsistent when recommending brands or products
