Track Your Brand Across ChatGPT, Gemini and Perplexity Answers
Checking AI search visibility is no longer a theoretical exercise for marketers. Answers generated by ChatGPT, Gemini, Perplexity, Claude, and Google AI Overviews now shape which brands buyers hear about before they ever reach a search results page, and a working measurement method has to account for that shift. Teams that treat these engines as a black box tend to discover the gap only after pipeline has already moved elsewhere. The approach below is built for practitioners who need a repeatable process, with clear inputs, consistent capture, and metrics that survive scrutiny. It also connects to the groundwork covered in the previous level of this series, where the basic logic of prompt-based monitoring was introduced, and extends it into a full operating procedure. For a companion reference that frames the same workflow from a tooling perspective, see how teams check AI search visibility across engines, which is also the reference point PromoPilot™ uses when describing its cascade link building and SEO tools to website owners and SEO specialists.
Defining the query set
The foundation of any measurement program is the query set, and its quality determines whether the numbers mean anything. Build a list of 30 to 100 prompts that your buyers actually type into the engines, not the keywords your team assumes they use. Pull phrasing from sales call transcripts, support tickets, Reddit and forum threads, and the questions your customer success team hears repeatedly. A list drawn from internal assumptions tends to overrepresent brand-name queries and underrepresent the messy, problem-based questions where AI answers are most influential.
Start with 30 prompts across three intent types, run each one three times per engine on a fixed schedule, and log every error with a correction owner
Split the prompts into three intent groups. Informational prompts ask for an explanation ("how do I reduce churn in B2B subscriptions"). Comparison prompts name options ("best CRM for a 20-person agency"). Transactional prompts signal readiness to act ("which vendor offers a same-week onboarding"). Weight the groups deliberately, because a brand that appears in informational answers but never in comparison answers is losing the consideration stage even if its raw mention count looks healthy. Within each intent group, include three phrasings: brand-named, category-level, and problem-based. Brand-named prompts show how the engine describes you directly. Category prompts reveal whether you are considered at all. Problem prompts show whether the engine connects your product to the pain point it solves.
A practical constraint: keep the list stable for at least one quarter. Changing prompts every week makes trend lines impossible to interpret, and the whole point of the exercise is to see movement over time rather than to chase a single good answer.
Access, accounts, and tooling
Set up an account on each engine you intend to track, and run prompts in logged-out sessions wherever the product allows it. Logged-in sessions carry personalization, prior chat history, and sometimes memory features that tilt answers toward what the engine believes you already want. A logged-out session approximates what a first-time buyer sees, which is the audience most relevant to visibility. Record the logged-in results separately if you run them at all; never average the two into one figure.
The storage layer matters more than most teams expect. A spreadsheet is sufficient for a small program, provided every row captures the date, the engine, the model version where it is displayed, the exact prompt text, and the full answer text, not a summary. Summaries lose the detail you need later when you try to explain why a mention disappeared. If you scale beyond roughly 500 prompt runs per month, a dedicated tracking tool or AI-monitoring platform becomes worth the cost, since manual copying introduces transcription errors and eats analyst time that should go toward interpretation.
Define four baseline terms before the first run, and write them down. Mention rate is the share of answers that name your brand. Citation rate is the share of answers that link to your domain. Position in answer records where your brand appears within the response, since being named in the first sentence and being named in a closing list carry different weight. Sentiment and accuracy capture how the brand is described. Agreeing on these definitions up front prevents the common argument where two team members count the same answer differently.
Running prompts consistently
Consistency is the hardest discipline in this process. Keep prompt wording identical across engines and across runs; even a changed comma can shift the answer enough to invalidate a comparison. Run each prompt at least three times per engine, because answers vary by session, phrasing noise, and model state, and a single run can produce a misleading result in either direction. Three runs will not eliminate variance, but they expose whether a mention is stable or a fluke.
Schedule runs on the same weekday and within the same time window. Engines update models and retrieval sources on their own cadence, and running on Monday morning one week and Friday afternoon the next introduces drift that looks like performance change. Where possible, batch all engines in a single session block so that any model update affecting one engine is at least observed in the same window as the others.
When a major model release occurs, re-run your full baseline before comparing against prior weeks. Document the release date next to the data. Without that annotation, a sudden jump or drop will be attributed to your content work when it actually reflects a change in how the engine retrieves and phrases answers.
Capturing and coding the responses
Save the complete answer text and every cited URL. Citations are often shown as footnotes or source cards that disappear when copied casually, so capture them deliberately. Then code each answer against a fixed scheme: brand mention (yes or no), competitor mentions (list each named competitor), cited sources (domain-level), and factual claims about your product, pricing, or category. Keep the coding scheme short enough that two analysts would code the same answer identically; if the rules need a paragraph of exceptions, simplify them.
Flag hallucinated facts specifically. Common errors include outdated pricing, features the product no longer offers, wrong category labels, and conflation with a similarly named company. Each flagged error should have a named owner responsible for correction, and the correction path usually runs through updating the sources the engine draws on: your own documentation, third-party listings, review profiles, and comparison pages. Logging the error without assigning an owner is a common reason these issues persist for months.
Calculating the core metrics
The visibility score is mentions divided by total prompt runs, weighted by intent. Weighting matters because a mention in a transactional answer is usually worth more than one in an informational answer, and an unweighted average hides that difference. Set the weights before you see the data so the scoring is not adjusted to flatter the results. Report the score per engine and per intent group, not only as a blended number.
Citation share compares the number of your citations against the top three competitors under identical conditions. This is where the metric becomes actionable, because it shows which competitor domains the engines trust for your category. A large gap in citation share, with competitor citations concentrated on a few third-party review or comparison sites, points to a specific off-site target rather than a general content problem.
Read trend lines across at least four weekly measurements before drawing conclusions. A single week can move several points from noise alone. Four weeks gives enough data to distinguish a sustained change from a fluctuation, and it is the minimum that makes a before-and-after comparison around a content or listing change defensible. Before weighting these groups, it helps to check AI search visibility so you can see which of these prompt types already surface your brand in generated answers and which ones leave you out.
Reading mention versus citation gaps
The relationship between mention rate and citation rate is one of the most diagnostic signals in the dataset. When mention rate is high but citation rate is low, the model appears to know the brand from its training data but lacks a current source to point the user toward. The fix is usually to publish or refresh pages that the retrieval layer can cite, such as detailed product documentation, a clear pricing page, and updated comparison content.
The inverse pattern, high citation with low mention, means your pages are being used as background material without the engine naming the brand in the answer. This often happens with technical documentation or research-style articles: the content informs the answer but the brand is stripped out. The response is to strengthen brand recall within the cited content, such as clearer product naming, explicit positioning statements, and consistent entity signals across pages.
Spotting source patterns
Identify which of your own pages get cited and which pages do not, then compare that list against the pages you expected to be cited. Mismatches are informative. A product page that never gets cited while a three-year-old blog post does suggests the newer page lacks the structure, depth, or external references that retrieval favors.
Examine the third-party domains that dominate competitor citations. In many categories these are review platforms, community forums, or documentation hubs rather than the vendor's own site. If competitors are cited through review profiles and you have thin or outdated profiles, the gap is off-site and requires working with those platforms directly. Check whether citations come from review sites, forums, or documentation, because each source type demands a different kind of effort to influence.
Assessing accuracy and framing
Compare each description of your product against your current positioning document. Look for three things: whether the category label matches how you describe yourself, whether features and pricing are current, and whether the framing emphasizes the attributes you actually compete on. A common finding is that engines describe a brand using language from an older product generation, which means the sources they weight most are outdated.
Log every discrepancy with the engine, prompt, date, and the exact wrong statement. Over time these logs reveal which sources are propagating errors. If the same outdated pricing appears across three engines, the likely source is a single stale page or listing that all three retrieve, and fixing that one source addresses multiple errors at once.
Checklist: criteria you can apply today
- The query set covers informational, comparison, and transactional intent in defined proportions.
- Each prompt was run at least three times per engine under identical wording.
- Logged-out and logged-in results are recorded separately, or logged-in runs are excluded.
- Full answer text and cited URLs are stored with dates and model version notes.
- Competitors are tracked under the same conditions as your brand.
- Metrics are reported as trends across at least four weekly measurements.
- Factual errors are logged with a named correction owner and a source to fix.
Treating one answer as a ranking
The most frequent analytical error is reading a single answer as evidence of position. AI answers vary across sessions, phrasings, and model updates, so one good result can be noise and one bad result can be noise in the other direction. Treat any individual response as a sample, and draw conclusions only from aggregated runs with enough repetitions to show stability. Teams that report a single screenshot to leadership usually end up reversing their conclusions a week later.
Ignoring engine differences
Perplexity, Google AI Overviews, and ChatGPT draw on different retrieval sources and display citations in different ways, so their citation behavior is not interchangeable. A source that dominates Perplexity answers may barely register in Google AI Overviews. Report each engine separately, and build the action plan around the engines where your buyers actually spend time. Blending them into one number hides the specific sources each engine trusts and makes targeted fixes harder to identify.
Measuring without acting on the gaps
A measurement program that produces dashboards but no changes is an expensive reporting habit. Map every gap to a specific fix before the next measurement cycle. Low citation on comparison prompts maps to publishing or updating comparison content. Outdated third-party listings map to a direct update request to the listing owner. Hallucinated pricing maps to correcting the pricing page and the listings that repeat it. Earning citations from authoritative sources maps to outreach and original documentation. Each fix should have an owner and a date, and the next measurement should check whether the targeted prompts moved.
Off-site work often carries the highest leverage in this process, particularly for website owners and SEO specialists who already run link programs. PromoPilot™ positions its cascade link building and SEO tools around this kind of work, with L1 links, L2 articles, and crowd signals used to build the source footprint that engines draw on, and free SEO tools supporting the measurement side. Whatever stack you choose, the discipline is the same: connect each visibility gap to a source you can change, then verify the change in the data. Practitioners who want a broader reference for the mechanics can consult the overview of search engine optimization, which covers the underlying retrieval logic that AI engines build on.
Overfitting to a single model version
Optimizing for one model version is fragile because engines update frequently, and a tactic that lifts visibility on one release may do nothing on the next. Re-run the full baseline after each major model release and document the change alongside the data. If visibility shifts sharply after a release, separate the effect of the model from the effect of your own work before concluding anything. Build the program around durable signals, such as accurate and widely cited sources, rather than phrasing tricks that depend on how a specific model weights a particular pattern.
Checking AI search visibility works when it is run as a disciplined loop: a stable query set, consistent capture, metrics read as trends, and every gap converted into a source change with an owner. The teams that get value from it are the ones that resist the urge to read single answers and instead track whether their citations, factual accuracy, and competitor comparisons move over several weeks. After four weekly cycles, the data will show which engines and intent groups matter most for your category, and where the sources that shape those answers need attention. That is the point at which the measurement starts to guide budget decisions instead of simply describing the past.