If ChatGPT names your brand, you have an observation about ChatGPT under recorded conditions. You do not yet have a conclusion about the market. I would use multiple AI experiences to locate where the buyer’s answer changes, then invest in the evidence behind the relevant gap.
Evidence map: compare like with like
| Source / design | Sample and period | What it measures | Interpretation limit |
|---|---|---|---|
| Semrush index | 126M US prompts; January–April 2026 | Industry-level source patterns | Not a paired sample of your questions |
| Academic preprint | 117 product questions; September 2026 | Displayed-domain overlap | English; Netherlands; preprint |
| Academic collection | 1,755 planned; 1,536 responses | Availability by condition | Missingness needs explicit handling |
| Your proposed panel | 10 questions × 4 surfaces × 3 repeats = 120 | Diagnostic of selected answers | Design only; not collected research |
The published benchmark shows meaningful differences
Semrush’s June 2026 announcement describes an analysis of 126 million US AI search prompts from January through April 2026 across ChatGPT, Gemini, Google AI Mode and Google AI Overviews, with benchmarks in 22 industries. It reports rounded averages of 15 cited sources per ChatGPT response and three per Gemini response. Only 36 brands maintained top-100 visibility across all four platforms in every month studied. Semrush, 2026 AI Visibility Index announcement, June 26, 2026
These findings support platform-specific measurement. They do not establish the result for your brand, category or Spanish audience. The announcement is a summary rather than a downloadable response-level dataset, so its averages cannot reconstruct a matched comparison for your own questions. My affiliation with Semrush is disclosed in the author biography; this article uses its public report.
Compare experiences, not interchangeable model names
A model, an assistant product and a search surface are different units. Record the product, visible model when available, mode and whether retrieval is enabled. AI Overviews and AI Mode are separate Google experiences; an AI Overview that does not appear is a different observation from a failed attempt to collect a result.
Google documents that those two experiences can use different models and techniques and therefore show different answers and links. ChatGPT may search automatically when a question benefits from current information, or a user can select search manually. A comparison that hides these conditions can mistake a mode difference for a brand problem. Google Search Central, AI features and your website OpenAI Help Center, Searching the web with ChatGPT
Choose surfaces using evidence about your audience and the decision. A four-surface panel is a useful starting proposal for a team investigating several experiences, not a universal requirement. A narrower documented buyer journey may justify a smaller panel.
A manageable panel still needs an honest workload
My proposed starting design is ten distinct unbranded buyer questions, four selected surfaces and three repetitions: 120 planned observations per round. That number is arithmetic, not a study sample already collected. It does not include identity checks that explicitly name the brand; keep those in a separate panel.
Cover definition, comparison, alternatives, case-specific suitability and selection. Use two genuinely different questions per intent if your decision requires all five. Write English and Spanish panels natively and analyze them separately; translating words does not establish equal demand or a comparable source ecosystem.
Version the question set before seeing results. Keep market, language and account context consistent, use fresh sessions under the same rules, and record the date and collection order. Save the complete answer and visible source URLs. Three repetitions help expose variability but are not three independent buyers. Collection and review take time; there is no defensible universal one-hour promise.
Score distinct signals with explicit denominators
Record a mention when the brand is named. Record a recommendation when it is presented as suitable for the stated situation. Record a citation when the answer links to a source, identifying whether it is your domain, your specific page or a third-party source. Review factual accuracy separately.
Calculate each rate over valid observations for that surface and round. Retain failed runs and explain exclusions. For AI Overviews, record valid searches without an overview as no overview shown; report trigger frequency separately. Do not count every non-trigger as an answer that excluded your brand.
Suppose, hypothetically, the brand is mentioned in 18 of 30 valid responses on one surface and six of 30 on another. Report 18/30 and 6/30 with the conditions. Pooling them as 24/60 would erase the difference you set out to diagnose. Do not treat either small sample as a market share estimate.
Turn a difference into a specific intervention
If an answer misstates your product conditions, inspect the sources supporting the error and correct the responsible public facts. If comparisons omit a relevant option, check whether your evidence explains fit and trade-offs. If your page is useful but never retrieved, inspect accessibility and eligibility before rewriting the entire library.
Assign an evidence owner and choose one intervention with a review date. Keep the question panel stable and document material product or mode changes. Improvement after an edit remains observational: new sources, platform updates, competitors and ordinary variability could also explain it.
You can preserve content when the evidence is accurate and no consequential gap appears. Expand the panel only when a material buyer question or experience is missing. The goal is a defensible diagnosis, not the largest possible dashboard.
The next move
Pick one buyer decision, document the relevant surfaces and run a versioned panel. Bring the differences, source evidence and material errors to the editorial review. Fix the gap you can substantiate before declaring that your brand is visible everywhere.
A paired audit shows why one screenshot is insufficient
A September 2026 academic preprint selected 117 manually validated physical-product questions from a pool of 2,528 commercial queries. It planned three repetitions across five conditions: ChatGPT interface and API, Gemini interface and API, and Google AI Overviews. Of 1,755 planned observations, 1,536 produced responses. Collection used predominantly English questions and a Netherlands IP. This is not a representative survey of buyers. arXiv, 2026 (preprint)
The paper reports mean displayed-domain Jaccard overlap of 5.4% between ChatGPT and Gemini, 5.2% between ChatGPT and AI Overviews, and 9.8% between Gemini and AI Overviews. Within-interface repetitions overlapped 26.0%, 29.8% and 45.9%, respectively. These measure source-set similarity, not brand citation probability or market share. arXiv, 2026 (preprint)
The pattern matters operationally: even repeating a question within one interface can change the visible source set. A single capture establishes what appeared once under those conditions. It cannot establish the stable answer a whole audience will see. Repetition improves the diagnosis while preserving uncertainty about broader behavior.
This is a preprint, not a claim of peer-reviewed validation. Its general product questions, market and collection dates restrict transfer to another category. The study is useful because it documents a paired design and its missing observations. It does not provide a league table of the best assistant for every business.
An API panel is not automatically a consumer-interface panel
The same study reports mean interface-versus-API domain overlap of 12.0% for ChatGPT and 14.8% for Gemini. It cannot establish identical underlying models across those conditions. Preserve that distinction when commissioning an automated visibility report: the convenience of collecting an API does not establish equivalence to what a buyer sees. arXiv, 2026 (preprint)
Ask the provider which product, model, retrieval setting, account context and location it observes. Ask whether it records displayed citations, retrieved sources or grounding metadata. These layers can contain different URLs. A retrieved URL does not establish that it was visibly cited or caused a recommendation.
If your monitoring uses an API, label the results accordingly and validate a relevant subset against the consumer interface. Keep interface and API series separate. An API can be a useful repeatable diagnostic without becoming an unsupported estimate of consumer exposure.
Make the overlap calculation reproducible
Jaccard overlap is the number of shared domains divided by the number of distinct domains in either answer. In a hypothetical pair, one answer cites alpha.example and beta.example; another cites beta.example and gamma.example. One shared domain divided by three distinct domains gives 33.3%. This example explains the metric and is not collected evidence.
Normalize consistently before calculating. Resolve a redirect where possible, remove tracking parameters and distinguish page-level comparison from registrable-domain comparison. Several URLs from the same publisher can count as multiple pages but one domain. Preserve the original visible URLs so a reviewer can inspect what normalization removed.
When both domain sets are empty, the denominator is zero. Report that case separately rather than assigning perfect agreement. When a response was not collected, mark missing. When a valid answer displays no sources, record an observed empty set. These states answer different questions.
Averages also need a defined aggregation rule. Comparing a mean source count in a broad industry study with a median in a product-question sample mixes populations and summaries. Differences in date, platform setting and what counts as a source may explain the figures. They are not interchangeable benchmarks for your brand.
Separate answer availability from brand performance
In the academic audit, Google produced 134 substantive, parsable AI Overviews across 351 searches, or 38.2%. That is an overview-availability measure in this sample. The remainder cannot all be labeled brand exclusion. First report whether the experience appeared, then assess the content of valid answers. arXiv, 2026 (preprint)
Use four statuses in the operational register: valid answer, valid search with no overview, collection failure, and missing evidence field. Inside a valid answer, separately record no brand mention and no visible citation. Keeping this hierarchy prevents technical failures from quietly lowering a reported brand rate.
For each question, preserve product and interface, visible model if known, market, language, collection time, repetition, status, full answer, visible URLs and the factual checks. Record unknown model information as unknown. A self-description in a generated answer is not independent verification of the serving model.
Summarize across questions, not as if repeated answers were independent people. If you calculate uncertainty, account for clustering by question and publish the method. Three repetitions of ten questions do not become thirty distinct buyer needs. Retain the per-question table so a strong aggregate cannot hide an important failure.
A correction workflow with observable exit criteria
Classify a gap before assigning work. A wrong product condition calls for a factual correction and a check of the supporting source. An unsupported comparison calls for selection criteria and evidence. A technically inaccessible page calls for an access review. These interventions require different owners.
Write an exit criterion that your team can observe: the public specification states the correct limit, the comparison documents an exception, or the relevant URL is accessible and eligible. Then repeat the fixed panel to see whether answers changed. Correct public evidence is a deliverable you can verify; universal recommendation is not.
Keep a change log covering the page, evidence revision, product updates and collection settings. If the answer improves, describe the association under those conditions and alternative explanations. If it does not, inspect the cited sources and remaining factual gaps before expanding monitoring. More observations only help when they answer an unresolved question.
Sources and scope
Public-source analysis, checked October 12, 2026. Proposed frameworks and examples are not findings from an original study. US data is not extrapolated to Spain.
FA