What does an AI answer actually tell you about your B2B brand?
Distinguish a mention, a citation and a recommendation
An answer may name your company, link to a page, or present you as a suitable supplier. Those are three different events. A brand mention shows the name appeared in one observed answer. A citation points to a source the reader can inspect. A recommendation makes a claim about fit. None proves that a buyer visited your site or became a qualified enquiry. The distinction prevents a visibility dashboard from turning every name drop into pipeline.
OpenAI says ChatGPT search can present links to sources. Google describes AI Overviews and AI Mode as Search features that can surface supporting links. Their answers and links are not identical products, so record the product and mode with each result. A screenshot of one ChatGPT answer cannot stand in for a cross-platform audit.
Start with what matters to your buyers. A category prompt such as “Which systems help a UK manufacturer manage supplier approvals?” asks a different question from “What does Acme's software do?” A B2B team should examine both discovery and verification. The AI search visibility service page explains the broader offer; this guide gives you a repeatable way to inspect a small sample yourself.
Do not confuse a sample with a universal ranking
Answers can differ when the wording, product mode, location context, available sources or date changes. You cannot know the share of all buyers' AI conversations from one manual test. Record the exact prompt and conditions, then compare like with like next time. If a tool produces a percentage, ask what prompts form the denominator and how often they were run. A high score on branded questions can hide absence from unbranded category questions.
An observation log should preserve the answer link or screenshot where permitted, the date, the platform, any visible sources and what was actually claimed. Keep a separate column for uncertainty: “name ambiguous,” “citation goes to directory,” or “answer uses an outdated product description.” This makes the audit a decision tool rather than a collection of favourable examples.
How do you run a practical AI search visibility audit?
1. Build a buyer-question set before opening the tools
Take questions from sales calls, proposals, implementation meetings and Search Console query themes. Group them by the decision a buyer is making: identify a problem, understand the category, shortlist vendors, compare options, check compatibility or validate risk. Include a few branded prompts to test factual accuracy, but keep the main set unbranded so the audit can reveal whether buyers could discover you without already knowing your name.
Write each prompt exactly as a buyer might ask it, without inserting your preferred keywords. Note the market or sector context only when it changes the decision. Freeze the first set before measuring; adding questions after seeing the results makes trend comparisons unreliable. The table below is a template, not actual search-volume or visibility data.
| Buyer stage | Example prompt | What to inspect | Relevant site evidence |
|---|---|---|---|
| Problem | How do procurement teams reduce supplier onboarding delays? | Is the problem described accurately? | Helpful process guide |
| Category | What software helps UK manufacturers manage supplier approvals? | Which categories and vendors appear? | Category or solution page |
| Fit | Which tools support approval workflows for multi-site teams? | Are requirements and limitations correct? | Use-case and integration pages |
| Comparison | How do these two platforms differ on implementation? | Are differences sourced and current? | Fair comparison and documentation |
| Risk | What security questions should we ask a vendor? | Which evidence is linked? | Approved security information |
| Brand | What does our company offer and who is it for? | Is the description accurate? | Product and company pages |
Example prompts are illustrative. Replace them with real customer language and record the exact wording used.
2. Capture a repeatable sample across the right products
Run the frozen prompt set in the AI products relevant to your buyers. Keep mode, date and language visible in the log. Save the answer, distinguish a named brand from a linked source, and record whether the statement about your company is true. For comparisons, note the competitors that appear without implying that every omitted company was evaluated and rejected. Repeat a small number of prompts to observe variation, but report the sample size rather than presenting an exact market share.
Do not scrape or automate a product in a way its terms prohibit. A manual sample can still reveal inaccurate descriptions and gaps in public information. If you use an AI visibility platform, inspect how it chooses prompts, geographies and variants. Its dashboard may be useful for direction, but its score remains a measurement of its own sample.
3. Inspect the source behind each material claim
When an answer cites a page, open it. Is the cited page yours, a review site, an old press release or a competitor's comparison? Does the source actually support the claim? A citation to an outdated directory may explain an obsolete category label; a current product page may still omit a limit that matters to buyers. Record source ownership and the specific correction needed.
For an uncited brand claim, look for repeated wording on your own site and other public listings. Correct what you control first: title and description, product scope, supported integrations, sector fit, pricing basis or contact details. Do not invent customer numbers or credentials to make the answer more persuasive. If a third-party page is wrong, ask its editor for a factual correction with evidence, not a guaranteed citation.
Which measures belong in the report?
Separate answer visibility from business outcomes
Use four columns. Observed mention rate is the number of frozen prompts in which the brand appears divided by the number tested. Observed citation rate counts prompts with a link to your site. Accuracy rate counts reviewed brand statements that match approved facts. Buyer-fit rate asks whether the appearance occurred in prompts relevant to the actual offer. Each is a sampled diagnostic with a documented denominator, not a market-wide ranking.
Then report visits and enquiries separately. Inspect ChatGPT referral traffic and the utm_source=chatgpt.com parameter when present in your analytics; OpenAI's ChatGPT search guidance explains how its search experience links to sources. A visit can happen without a tracked conversion; an enquiry can happen after several channels. The B2B SEO measurement page covers the CRM hand-off. Avoid assigning a later deal wholly to an AI mention you happened to observe.
Google's Search Console generative AI reports provide a dedicated view of impressions and pages in eligible AI features, while Google also includes those appearances in overall Search performance. Microsoft's Bing Webmaster Tools AI Performance preview reports cited pages and citation activity across supported Microsoft experiences. These platforms measure different surfaces; do not add their counts together as a single “AI reach” figure.
Use a decision log, not a vanity-score chart
For each missed or wrong answer, write a hypothesis, owner, source page, change and recheck date. “No mention on category prompts” may point to unclear solution pages or thin independent evidence. “Wrong integration listed” points to documentation accuracy. “Cited but no relevant visits” may be normal for an answer that resolved the question, or it may indicate a weak next step on the cited page. Investigate before choosing a fix.
Prioritise errors by buyer risk. An inaccurate security or capability claim deserves faster correction than a missing link on a broad educational prompt. A useful report can have a low mention rate and still identify one serious fact to fix. The SEO reporting guide shows how to keep diagnostic indicators separate from qualified demand.
The zero-click B2B search guide explains why a search impression, citation, visit and qualified enquiry cannot share one total.
What can you change on the site without chasing a secret algorithm?
Fix access and answer quality first
Check whether important public pages are crawlable, indexable and easy to find through internal links. OpenAI distinguishes OAI-SearchBot from other crawlers, including GPTBot; review the current user-agent policy rather than copying a generic “block all AI” rule. Google's AI features guidance says eligible supporting links must meet normal Search requirements and that no special AI markup is required. Neither provider promises a citation when a page meets those requirements.
Make the page itself a better source: state who the product serves, what it does, prerequisites, limits and the date of any volatile claim. Use named expertise where genuine, link to detailed implementation information, and correct stale third-party listings. Structured data should match visible text. A hidden “LLM optimisation” file cannot replace a clear product page and trustworthy source trail.
Review the same prompt set after meaningful changes
Recheck the frozen prompts after a material page correction or at a regular interval. Compare mentions, citations and accuracy with the earlier sample, but note product changes and sample limits. Save the previous answer alongside the new one. If the answer remains wrong, check whether the cited source has changed and whether the corrected page can be accessed; do not assume instant recrawling or a direct cause-and-effect relationship.
The strongest result is a buyer getting a correct description and a useful route to verify it. If you need a wider audit of search accessibility, source quality and enquiry tracking, book a B2B SEO audit or review the B2B SEO services that support the work.
Want this applied to your own site?
The free audit is a 20-minute review that ends with three specific opportunities.
Use the ChatGPT citation guide for source quality and the crawler policy guide for access checks.