If you sell more than one product, checking whether ChatGPT knows your company is not enough. It might name the wrong edition, describe a feature the product does not have, or send people to your homepage when it should link the product page. This guide is for marketing leads who want to track the product level mentions ChatGPT gives each of their products, and end up with a report backed by saved answers.
The short version: define the exact product names you will count, match buyer questions to each product, and rerun them under the same conditions. Save every answer, separate a mention from a recommendation and from a page citation, and calculate results for each product on its own.
What you need: a list of the products, models or editions you want to watch, the current product pages and documentation for each, access to ChatGPT Search, and a spreadsheet. DeepSmith is optional and helps in three of the steps below, as one AEO tool product mentions workflow among others. For a wider look at the category, see our guide to AI brand monitoring tools.
Step 1. Define the exact product you will count
Before you run any prompt, write down what counts as a hit. Create one record for every product, model or edition you want to monitor, because "the product" is rarely one thing. Each record should hold a stable product ID, the canonical name, approved aliases, excluded matches (similar names, sibling products, generic uses of the word, wrong model numbers), the family and variant, the primary product page, any other approved destinations, the buyer use cases, the market and language, a current fact sheet with a last-checked date, and an owner who settles identity and accuracy questions.
Here is a made-up example so you can see how this works. Say the product is Northstar Signal Board, and a separate edition called Northstar Signal Board Lite is monitored on its own. "Signal Board" is an accepted shorthand, but only when the answer clearly points to the Northstar product. The word "Signal" on its own is ambiguous. If an answer only names the family, you record a family-level observation and do not award an exact-edition mention.
Done when: a second person can use the registry on a few sample answers and sort each one into a confirmed match, an excluded match or an ambiguous match.
Common mistake: searching for a short product word and counting every time it appears. Matching the exact text is a good way to find candidates, but it will not settle common words, similar names, editions or model numbers.
Step 2. Map buyer prompts to each product
Give every product its own fixed set of buyer questions, and split them into two groups right away, because they answer different questions.
- Discovery prompts describe the buyer's job, constraints or category without using your product's name. They tell you whether ChatGPT brings the product up for a real need.
- Named-product diagnostics do use the product's name. They tell you how ChatGPT describes the product, compares it and finds its page, but they say nothing about whether ChatGPT would find it unprompted.
Mixing the two flatters your numbers. The guide on branded, unbranded and competitor prompts goes deeper. These templates are a good start for each product:
| Prompt purpose | Template | What to look at |
|---|---|---|
| Category discovery | Which products help [buyer] accomplish [job]? | Whether your product appears at all |
| Constraint fit | What should [buyer] use for [job] when [constraint]? | Whether it shows up where it is suitable |
| Alternative discovery | What alternatives to [competing product] fit [specific need]? | Whether it appears in a competitive buying context |
| Named-product fit | Is [exact product] suitable for [use case]? What are its limitations? | Accurate positioning and limits |
| Named-product facts | What does [exact product and edition] include, and where is its official product page? | Identity, accuracy and the right destination |
Only include a requirement your product actually meets, or an absence tells you nothing. For the made-up product, a discovery prompt might be "Which tools can a small product team use to collect and organize customer feature requests?" and a diagnostic might be "Is Northstar Signal Board the same product as Northstar Signal Board Lite?"
For a first pilot, a good size is six discovery prompts and two named diagnostics per product, each run in three separate collection windows. That gives you 18 discovery answers and six diagnostic answers for a product. This is a manageable starting point, not a proven minimum, so add prompts when a product has very different use cases, editions or markets. Setting up a repeatable prompt set covers the routine side of this.
Where DeepSmith fits: Discover Prompts suggests questions based on your product and persona context, which gives you a starting list. Check each suggestion by hand to be sure it belongs to the product and the prompt group you have in mind. A suggested prompt does not prove demand, and it does not create a product-level report on its own.
Done when: every prompt has a stable ID, exact wording, a product it belongs to, a discovery or diagnostic label, and a reason the product is relevant to it.
Common mistake: asking ChatGPT to include your product and then reporting its appearance as evidence of discovery.
Step 3. Fix the ChatGPT test conditions
Answers change with the setup, so pick one and write it down. Run each standalone test in a fresh conversation. Earlier answers in a long chat can hand product names to later ones. If you want to test a multi-turn buying journey, store it as its own test.
OpenAI's current instructions for an unpersonalized Temporary Chat are to open a new chat, select Temporary, choose Unpersonalized before the first message, and then enter your prompt. Unpersonalized chats do not use memory, custom instructions or plugins. You cannot change the choice after the conversation starts, and not every temporary chat turns personalization off by itself, so pick it on purpose.
For chatgpt monitoring for product pages, make ChatGPT Search your main test condition so you can inspect the source links. ChatGPT can also search on its own, so write down what actually happened, and keep any non-search answers as a separate group.
For every run, record the timestamp and timezone, the interface, the account plan, the model or mode label if it is visible (write "unavailable" if not), the Temporary Chat personalization choice, whether Search was selected and what it did, the intended buyer country and language, the location-sharing state, the exact prompt version, and whether it was a fresh conversation or a documented sequence.
OpenAI says Search may use your IP address to estimate a general location, and device location sharing is off by default. Asking for "buyers in Canada" states your intent but does not control every location signal, so log both the market you meant and the conditions you actually had.
Done when: another team member can rerun the same prompt with the same recorded setup, knowing that an identical answer is not guaranteed.
Pro tip: consistent conditions make comparisons more useful, but they do not make ChatGPT give the same answer twice.
Step 4. Collect a baseline and save every answer
Now run the fixed prompt set across your planned collection windows and save the evidence before you score anything. For each run, keep the run, product and prompt IDs, the full prompt text, the full answer (not only the sentence with the product in it), the timestamp and conditions, the visible inline citations and where they point, what the Sources panel shows, any product cards, and the run status: completed, clarification, refusal, technical failure or access limit.
Temporary chats do not show up in normal history, and OpenAI says an unsaved one cannot be reopened after you leave it. Copy the answer out before you navigate away. Archiving gets its own walkthrough in how to capture and archive verbatim AI answers.
For a manual pilot, run the same prompts in three separate windows, then pick a schedule you can keep up and review the report weekly. Daily collection can make sense if product facts change often, but it is a choice and not a rule. Do not regenerate until you get a favorable answer. If retries are allowed, set the rule first and keep every attempt.
Where DeepSmith fits: DeepSmith runs your tracked prompts on a schedule, captures the full answer, and keeps an answer history you can open. Raw visibility data can be exported as CSV or through the API. Use it to collect the product-mapped prompt set and read the real answers. If you are weighing this against doing it by hand, our comparison of an AI visibility tool and manual monitoring lays out the tradeoffs.
Be clear about the limits. A product-specific prompt does not make a standard mention metric product-specific, so the links between prompts and products, and the product calculations, may need to live in your review table. A DeepSmith run and a manual session may not use identical conditions, and shopping cards may not be captured, so label your method and mark card fields unavailable if so.
For an analytics team, DeepSmith's API is read-only. Workspace owners create keys in Settings under API keys, and the key is shown once. The visibility read scope covers answers, cited pages and exports, with a limit of 120 requests per minute per workspace. Rates come back as fractions between zero and one, so 0.42 means 42%, and full-answer retrieval returns up to 200 answers, newest first, so check a large report for completeness.
Done when: your collection log shows expected runs, completed runs, failed runs and the evidence for each answer, and any missing or capped retrievals are visible.
Common mistake: keeping only the answers that name your product. That throws away the denominator.
Step 5. Classify the exact product in each answer
Next, apply your registry and a fixed rubric to every completed answer, including those where the product is absent.
For each answer, record a confirmed textual mention (1 or 0), the match type (exact name, approved alias, family-only, wrong variant, ambiguous or absent), whether it was recommended for the stated need, product-card presence (present, absent or not captured), an accuracy status, the product-page citation status, any other product-specific citation, the exact excerpt behind your call, and who reviewed it.
Count a product once per answer for the main mention metric, even if the name appears several times. Word frequency is not the same as presence in a response. A mention and a recommendation are different, and our piece on telling citing from mentioning covers that split in more detail. Here is how a few typical answers score for the made-up product:
| Answer behavior | Mention | Recommendation | Treatment |
|---|---|---|---|
| "Northstar Signal Board is one option for this workflow" | 1 | 1 if genuinely proposed as suitable | Save the evidence and check the suitability claim |
| "Northstar Signal Board does not meet that requirement" | 1 | 0 | Presence is not endorsement |
| Names the parent company but no product | 0 | 0 | Do not infer a product mention |
| Names Signal Board Lite when you monitor Signal Board | 0 | 0 | Record the wrong variant separately |
| Lists the product only in the Sources panel | 0 | 0 unless the answer recommends it | Evaluate the source record separately |
When identity stays unclear, give the confirmed mention a zero and keep an ambiguity flag, then work through that queue before the final report. Do not quietly drop the hard rows. If you use a model to flag matches, OpenAI's evaluation guidance favors explicit criteria and pass or fail judgments, and a person should still check common names and variants.
Decide your denominator before you report anything. A valid run is a completed answer under the planned conditions, and an answer that names no product is still a valid zero. A clarification or refusal is kept, flagged and counted as no confirmed mention. Technical failures and access-limit errors are collection failures, not product absences. Never discard unfavorable answers after seeing the result.
For one product and one fixed discovery group, your textual mention rate is the valid answers with a confirmed mention divided by all valid answers. The recommendation rate and the product-page citation rate use the same denominator, with the numerator being answers that recommend the product, or that cite the approved primary page.
Calculate card presence separately, and report it as unavailable, not zero, if your method does not capture cards. Give the named-product diagnostics their own results, and never fold them into the discovery numbers. If one answer mentions two of your products, score each on its own.
Here is a worked example with invented numbers, not a benchmark or a customer result. Say a product has 20 valid discovery answers. Seven contain a confirmed mention, which is 35%. Three cite the approved product page, which is 15%. The last comparable window had five mentions out of 20, which is 25%, so the mention rate moved 10 percentage points. Show counts next to percentages, because a swing of two answers is not proof of a market shift.
Done when: every scored row has evidence, a fixed rule and a product identity, and you could recalculate the report from the saved answers.
Common mistake: treating the number of times a name appears, the number of answers, and the number of page citations as if they were the same count.
Step 6. Check that ChatGPT described the right product
A correct name does not mean a correct description, so compare each answer's claims with your current fact sheet. Go through the claims in the same order each time: identity (product, edition, generation, market), core use case, capabilities and limits by edition, compatibility, commercial facts like price and availability, and whether the recommendation respects the buyer's constraints.
For each disputed claim, write down the exact wording, the authoritative fact, the date you checked it, and the likely consequence. Use "not assessable" when the evidence does not settle it. A simple severity scale is critical for a wrong product, edition or made-up capability, material for outdated commercial terms, and minor for loose wording. A product can earn a mention and still fail the accuracy check, so keep both. If a description is wrong, the guide on how to audit and fix what AI says about product features covers what to do next.
Shopping answers need extra care. OpenAI says shopping intent can produce product options with images and links, chosen independently of ads, and not every available product will appear. ChatGPT may simplify titles using third-party information, labels like "most popular" are not verified guarantees, and the first price shown usually reflects the first merchant listed.
Done when: each material product claim is checked or marked unknown, and presence is not confused with accuracy.
Common mistake: reporting a generated "most popular" label as measured popularity, or treating a merchant's displayed price as the product's current price everywhere.
Step 7. Audit product-page citations separately
Now check where ChatGPT links, and whether that is the right page for the product, which is the core of chatgpt monitoring for product pages. In Search answers, OpenAI says responses may include inline citations. On desktop web you can hover over one for more detail and click to open the source. If no inline citations show, the Sources link under the answer opens a panel with cited sources and other relevant links. Do not count every item in Sources as a confirmed citation unless the interface shows that relationship, and mark uncertain ones as unconfirmed. If the line between a citation and a mention is still fuzzy, what counts as an AI citation is a good reference.
Sort each destination into a type: the exact primary product page, a correct regional or edition-specific page, documentation or specifications, a third-party review or comparison, a homepage or category page, a merchant purchase offer, a wrong product or variant, or a broken or unresolved destination.
Your main product-page metric should only use the destinations you named in the registry. Do not combine documentation, homepages, reviews and purchase links into one success count, and save the original destination before following redirects.
Where DeepSmith fits: the Pages view shows which of your pages AI cites, each page's share of your citations, and the prompts driving them. Use it to look into an approved product page, then go back to the matching ChatGPT answers to see the context. Page-level attribution does not show that the exact product name appeared correctly in the answer. DeepSmith does not map pages to products on its own or alert at the SKU level, so keep that mapping in your registry.
If the product is mentioned but the primary page is not cited, look at which sources support the description. If the page is cited but the product is not named, do not award a text mention. If the wrong edition page is cited, check edition naming and your registry. If there is no mention and no citation, recheck prompt fit and conditions before assuming a content problem.
When a product page is missing from citations, check that it is publicly accessible and that the name, edition, capabilities and limits are clear on it. OpenAI separates OAI-SearchBot, which supports ChatGPT search discovery, from GPTBot, which is about possible model training, and a site can allow one and disallow the other. OpenAI recommends allowing OAI-SearchBot in robots.txt and permitting its published IP ranges, and says systems can take about 24 hours to adjust after a robots.txt change. Our explainer on how ChatGPT fetches and renders your pages goes further. Access is a condition for being picked, not a promise of citation.
Done when: every product-page citation you claim has a verified destination, a source relationship, a prompt and a saved answer.
Common mistake: calling an ordinary purchase link a source citation, or counting a homepage citation as proof that ChatGPT cited the product page.
Step 8. Compare like for like and pick the next action
Report each product using the same prompt group, rules and conditions in every window. Your aeo tool product mentions report should show the exact product and edition, the test scope, coverage (expected runs, valid answers, technical failures, unresolved identity cases), mention, recommendation and product-page citation counts and rates, accuracy issues with severity, prompt-level gaps, saved evidence, and an owner and retest date for each action.
If you change the questions, product names, models, modes, markets or collection method, mark a break in the series or start a separate one. Otherwise you will credit a visibility change to what was only a change in your test. Broader AI visibility tools report at the brand level, which is a different job from this one.
Then turn findings into an action queue. A misidentified product means making naming and edition differences clearer in your own material, then retesting. An incorrect capability means verifying the fact and correcting your pages if needed. A wrong destination means looking at page identity and redirects. An absence for a relevant question means checking whether the product is clearly documented for that need, because absence alone does not say why it was left out. Weak evidence or missing runs get fixed before you make any marketing decision.
Choose your alert rules before you look at the outcome. For example, repeated absence across comparable windows can trigger an investigation. That is an operating rule and not a validated threshold, and alert thresholds for AI visibility changes are worth working out before you rely on one. There is no universal good mention rate, so judge each product against its own baseline. A page update may be followed by better answers without having caused them.
Done when: the report can be reproduced from saved evidence, and every meaningful issue has an owner and a retest date.
Common mistake: publishing a single composite score that hides identity errors, missing data, prompt changes and accuracy problems.
What to do next
Start with one product. Set up its identity record and prompt set, and collect a comparable baseline before you add the rest of the catalog, so the product level mentions ChatGPT returns are easy to compare. When you start reading the results, fix incorrect descriptions and wrong-product matches first, since those do more harm than a lower percentage. These results are samples of how ChatGPT answered your questions, not a count of people who saw your product.
If you want recurring collection without running every prompt by hand, you can use DeepSmith to gather ChatGPT answers and look at the pages they cite, then apply your own product checks to the saved evidence. Start a free trial to build your first baseline.
Related guides
- xFunnel Alternatives: 8 Best AI Answer Analytics Tools for 2026: Ranks eight xFunnel alternatives (DeepSmith, AthenaHQ, Profound, Rankscale, LLMrefs, OtterlyAI, Peec AI, Scrunch) against eight tests including answer visibility measurement, granularity, citation intelligence, competitive benchmarking, actionability, governance and.
- LLM Monitoring Explained: Brand Visibility Tracking vs Internal Usage Monitoring: Tell apart LLM monitoring for brand visibility and LLM monitoring for internal app usage, so you search for the right tools.
- Why Monitor AI Answers for Category-Level Narratives, Not Just Your Brand: Learn why watching how AI describes your category matters, and what to log beyond your own brand.
- How to Monitor Your Brand on Perplexity, Grok, Meta AI and DeepSeek: Decide which of the secondary engines are worth monitoring and how to set each up.
- What Is AI Brand Monitoring and How Does It Work?: Understand what AI brand monitoring watches, how it differs from simply checking an answer, and what outputs it produces.
