If you need to track multiple products AI assistants recommend, you have probably noticed that checking ChatGPT or Perplexity by hand stops working fast. You can't tell which products are showing up, which ones lost ground, or what to tell leadership on Monday. This guide walks you through eight steps to track multiple products across AI assistants, keep each product's evidence separate, and roll everything into one leadership view that still lets someone click down to the actual answers. It's written for marketing leads who run one tracking program that covers many products, including a multi brand retailer AI visibility program, the kind that enterprise AI visibility tools are built for, where the catalog is large and the questions come from several directions.
One thing to know before you start: a question that is about a product does not automatically give you a product-level number. A tracker reports brand and prompt level results, so the link between a product and its prompts, and the rules for counting an answer as a product win, are things you keep in your own registry or spreadsheet. That sounds like extra work, but it's mostly a table you fill in once and update now and then.
Step 1: Define what counts as a product
Start by deciding what one "product" means in your report. It might be a product family, a single sellable item, or an individual variant such as a size or color. Pick the level you actually want to report on, and keep the other levels separate so they don't blur together.
For a retail catalog, a stable internal product ID can identify the parent product, and separate variant IDs can cover each sellable variant. Google's guidance on product variants draws the same line between a shared group for the variants and a distinct item for each one, and each variant has its own GTIN. If you use GTINs as an extra identifier, use only values you have verified, and never guess one.
Create a registry with these columns, at a minimum:
- A stable internal product ID and the canonical product name.
- A parent product or family ID, if it applies.
- A variant or SKU ID, if you are measuring at variant level.
- Approved aliases, abbreviations, and names that AI answers might use.
- The preferred product page, if there is one.
- A written rule for whether a result about a family counts for each of its variants.
Treat the alias list as a matching aid and not as proof. Similar names, generic words, bundles, and model numbers all create false positives, so have a person look at ambiguous matches. A product that is only implied, or one that gets confused with another product, should not be counted as a confirmed mention.
Done when: every reporting unit has one stable key, a canonical name, and a clear rule for its variants and aliases.
Common mistake: treating a product family and all of its variants as the same thing. A family-level mention can then look like proof that one particular SKU was recommended, and it wasn't.
Step 2: Build a prompt registry and map prompts to products
Next, keep one record for each exact question you track. Anchor each record to the platform's prompt ID where there is one, and keep the wording exact. A shortened version of a question is a bad join key, because a similar-looking question can have a different intent, and you want to look at it before you call it a duplicate.
Useful fields for each prompt:
- The prompt ID and the exact prompt text.
- The product ID or IDs the prompt is meant to measure.
- The prompt set name and version, with the date it was frozen.
- One primary topic, buyer stage, and persona, as separate fields.
- The intent, such as product discovery, comparison, or use-case advice.
- Region and language, where they matter.
- Active or paused status, kept apart from buyer stage.
Some questions compare several products on purpose. Mark those as a comparison group, or map them to each product under a rule you write down. Please don't quietly give one answer to every product named in the question. It also helps to keep broad category questions apart from named-product questions, so your dashboard can show category discovery on one side and named-product visibility on the other. If you want a longer walkthrough of tagging, organizing tracked prompts by topic, stage and persona covers it.
DeepSmith's Discover Prompts can suggest candidate questions from your product, persona, and buyer-stage context. You review each suggestion and decide which product and group it belongs to. A suggested prompt isn't evidence of search demand, and it doesn't create product-level reporting on its own. DeepSmith's prompt categories give you category-level mention and citation rates, including a row for uncategorized prompts, and you can use those as one of your product groupings if you set them up that way.
Done when: every active prompt has a stable ID, exact wording, a version, and either a product mapping or an explicit category or comparison label.
Common mistake: joining records by a paraphrase of the question and then losing or misattributing the answer history.
Step 3: Create comparable prompt sets for each product
Now build a balanced set of buyer questions for each product. Tracking only the brand-name query tells you very little, because most buyers ask about use cases, comparisons, and options before they ever type a product name.
Use the same families of questions across products wherever they make sense: discovery, use cases, comparisons, and evaluation. Then add product-specific questions where your products truly differ. Tag each prompt with one primary topic, one buyer stage, and one persona, and combine those dimensions when you analyze, instead of squeezing them into a single label that means three things at once.
Keep a stable baseline set. If you add, remove, or heavily rewrite questions, make a new version of the set and note the change in your reporting, because a run on one version doesn't compare cleanly with a run on another. Keep the full answer history too, so an analyst can look at what changed rather than relying on a summary rate. If you haven't built a baseline before, setting up a repeatable prompt set shows the basic structure.
Pro tip: show the prompt count next to each rate. A product with no prompts in a category has a measurement gap, which is different from having zero visibility.
Done when: each product set has documented coverage of the buyer questions you care about, and you can compare the recurring baseline over time without the wording or membership shifting underneath you.
Step 4: Lock the collection conditions and schedule
Before you compare products with each other, write down the conditions that can change an answer: the engine, the exact prompt version, the collection period, the region and language, and the account state. A logged-in session and an anonymous one can behave differently because of personalization and memory. Keep these conditions steady for your baseline, and when something has to change, label it so nobody reads the difference as a trend.
DeepSmith runs your tracked prompts on a schedule you set, captures the answers, and keeps the answer history. Its visibility views support 7-day, 30-day, and 90-day windows, and the tool reference documents 30 days as the default. Rates in the data are fractions between zero and one, so 0.42 means 42%. The views also return the previous window's value and the change next to the current rate, which gives you a ready comparison. Use the same window, prompt version, engine coverage, and product-mapping rules whenever you report a product's change.
DeepSmith tracks ten engines: ChatGPT, Gemini, Perplexity, Claude, Google AI Overviews, Google AI Mode, Grok, Meta AI, Microsoft Copilot, and DeepSeek. Which ones you get depends on the plan. Pro tracks ChatGPT, Grow adds Perplexity, Scale adds Gemini, and Enterprise and Custom cover all ten. If you're still choosing a platform, AI visibility tools compares the main options. Tracked prompt limits are 50 for Pro, 100 for Grow, and 200 for Scale, with Enterprise set to your needs. Those limits belong to the workspace, so they are not a separate allowance for each product. If you have 20 products and 100 prompts, you are dividing 100 prompts across 20 products, and it's worth doing that arithmetic before you promise coverage. If you aren't sure which engines to start with, pick the ones your buyers most often use for product questions.
Done when: the schedule, plan coverage, reporting window, prompt version, geography, and account state are written down and applied the same way every time.
Common mistake: presenting a change in prompts, locale, engine mix, or personalization as if the product gained or lost visibility.
Step 5: Collect answers and label each outcome separately
Be clear about what one observation is. One observation is one answer, for one prompt, on one engine, at one collection time. For each one, keep the exact prompt, the answer text, the timestamp, the engine, the cited URLs, and whatever evidence you used to classify it. It also helps to record where in the answer the product appears, whether an owned page was cited and which one, which competitors showed up, and any operator notes about outages or model changes.
For each answer you can fairly count, label these outcomes separately:
- Text mention: does the answer name the product, including a verified alias?
- Recommendation: does it actually recommend or select the product, or only mention it?
- Product-page citation: does it cite the approved product page or another owned page that qualifies?
- Shopping-card presence: was the product in a card or carousel, if your collection method captures that surface?
- Accuracy: is the product's identity and description correct?
One answer can earn more than one label. A product-page citation doesn't prove the answer named the product, and a mention doesn't prove the product was recommended or described correctly. If an answer links to the same product several times, count it once in an answer-level rate. If your method can't capture shopping cards, write "unavailable" and leave out the zero.
This matters most for AI shopping visibility reporting, because a shopping answer is several things at once. ChatGPT's help documentation says product results are chosen independently and aren't ads, and it also says the product carousel reflects the model's reading of the user's intent and context, which can include Memory or Custom Instructions. Not every available product appears, and ChatGPT may simplify titles or write its own labels and review summaries, so those generated details shouldn't be treated as verified product facts. The product it picks and the list of merchants it shows afterward are two separate observations. Perplexity's Instant Buy help page says listings are ranked on criteria like source authority and relevance, and that merchants with deeper product details are more likely to be recommended. For Google AI Mode, Google describes shopping answers as combining product visuals with price, reviews, and inventory. For the other engines, test only the surfaces your plan and method really cover, and don't assume shopping cards or merchant lists exist where you haven't seen them. If you want the engine-by-engine detail, tracking visibility in AI shopping results goes deeper on these.
DeepSmith's AI Visibility gives you scheduled answers and answer history, per-prompt rates, platform and category views, citations, and cited-page information. The Pages view shows which of your pages were cited and which prompts drove those citations. A page-level citation doesn't establish that an exact product or SKU was named correctly, so the product-level labels still come from a review step you define, whether that's a person, a script, or both. The step-by-step version for one engine is in monitoring product-level mentions in ChatGPT.
Done when: every product outcome you count traces back to a specific answer and prompt, and anything uncertain or uncaptured is marked that way.
Step 6: Calculate rates with explicit denominators
For a fixed product, prompt group, engine, and time window, calculate each answer-level rate using the valid answers you observed as the denominator. Write the definition in the report so nobody has to guess.
- Product mention rate: valid answers with a confirmed product mention, divided by eligible valid answers.
- Product recommendation rate: valid answers that recommend the product, divided by the same eligible base.
- Owned product-page citation rate: valid answers that cite at least one approved owned product page, divided by the same base.
- Shopping-card presence rate: valid captured answers that show the product card, divided by valid answers where card capture was available. Never substitute zero for unavailable capture.
- Accuracy rate: answers that pass your product-identity and factual review, divided by the answers you reviewed.
These are suggested definitions for a custom report. They are not DeepSmith's own rate denominators, and DeepSmith returns its own platform-level and prompt and category rates. Keep those as they come, rather than recalculating them from an incomplete set of retrieved answers. Next to each custom rate, show the number of distinct prompts and the number of eligible answers.
When you need one portfolio number, say how you built it. An answer-weighted total is all qualifying answers divided by all eligible answers across the products included, so products with more observations count for more. An equal-product average gives every product's rate the same weight. They answer different questions, and neither is a universal portfolio score. Keep the per-product results visible beside the summary, and say how many products, prompts, and answers went into it.
For period-over-period change, compare matched prompt sets and windows. If a product's prompt mix changed, call the result a mix change, or use a stable baseline group for the like-for-like comparison. A segment with no prompts is not a zero. And there is no universal "good" mention rate in the sources we reviewed, so judge each product against its own baseline and its own coverage.
Done when: every rate in the report has a definition, a denominator, a prompt count, and an answer count next to it.
Step 7: Build a leadership view that keeps the drilldown
Put the portfolio summary at the top, then give people a route down to the product, engine, category, prompt, answer, and cited page. A leadership table that works well has these fields:
| Field | What it tells the reader |
|---|---|
| Product name and stable product ID | Which unit is measured, and no family or SKU confusion |
| Engine and window | The scope of the result |
| Metric and current value | Mention, recommendation, citation, card presence, or accuracy, kept apart |
| Prior value and change | The direction against the matched comparison period |
| Prompt count and eligible answer count | How much coverage sits behind the number |
| Prompt set version or group | Whether the comparison is like for like |
| Evidence link or record key | The way back to the answer, the cited URLs, and the label |
| Coverage or data-quality note | Missing prompts, unavailable cards, changed conditions, or incomplete retrieval |
A matrix with products as rows and engines as columns is a good way to see gains and losses at a glance. Don't make it the only thing you hand over, though. A dashboard built for AI visibility can hold both. Add a prompt and answer drilldown so a leader can see why a product moved, and show the gain or loss for each product along with how the portfolio number was calculated. Keep competitor mentions, your owned citations, and competitor citations apart from each other. For ideas on what leaders tend to want on the top page, see an AEO reporting dashboard for marketing leads.
DeepSmith's overview reports workspace-wide mention rate, citation rate, share of voice, trends, a platform breakdown, a competitor leaderboard, and cited sources, so it makes a good first layer for this view. The product-specific mapping and rollup live in a spreadsheet or BI model of your own. The data comes out through CSV export or the read-only API. The API uses scoped keys (visibility:read for visibility data), allows 120 requests per minute per workspace, and reads don't spend credits. It covers workspace overview and trends, platform and category breakdowns, the competitor leaderboard, prompts and prompt analytics, cited pages and sources, and the answers themselves. The suggested data model in the guide to using an AEO platform API for custom reporting is an example, and its column names are not literal API fields.

A simple structure for your own reporting tables is five tables: products, prompts, answers, outcomes, and a snapshot table where you save the rates DeepSmith returned, apart from your own calculations. Keep original IDs and timestamps at full precision, save raw API responses before you transform them, and build joins and de-duplication so scheduled refreshes don't create duplicate answers.
Google Search Console has a Generative AI performance report, and it makes a useful extra view for Google. It reports impressions for links to your site shown in Google AI Overviews and AI Mode, with page, country, device, and date breakdowns, and you can read how it works in the Search Console help page. It counts impressions of links, which is a different thing from the product mention and recommendation rates above, so don't add the impression count to answer-level rates as though they shared a denominator. Newest data can be preliminary, and chart and table totals can differ.
Done when: a leader can see which products gained, which lost, and which don't have enough measurement yet, and can open the prompts and answers behind each result.
Step 8: Audit coverage and plan the next collection cycle
Before you share anything, run through a short checklist. Does each product have the active prompts you meant it to have? Does each prompt have an ID and a current version? Is every engine and window labeled, and does every rate have its denominator? Look for duplicate prompt IDs, missing answers, answers you haven't classified, alias collisions, uncaptured cards, and changes in locale or account state. Also check that your answer retrieval is complete. The documented full-answer tool returns up to 200 answers, newest first, so a larger review needs pagination or an export, and you shouldn't assume the first set of results covers the whole period.
AI answers vary, and asking the same question at different times can give different answers and citations. One 2026 preprint, Quantifying Uncertainty in AI Visibility, looked at Perplexity Search, OpenAI SearchGPT, and Google Gemini across three consumer-product topics. It collected daily over nine days and sampled every ten minutes, and it reports substantial variability in citations. It cautions that a single run can look more precise than the evidence supports. A separate paper, Don't Measure Once, studied four engines and Swiss-German campaigns, and found limited day-to-day overlap in cited sources and brands. Its region, language, prompts, and verticals limit how far you can generalize, so use both as a reason to sample more than once and not as expected numbers for your own catalog.
DeepSmith's guidance on repeatable prompt sets suggests running each prompt at least three times per engine per collection window and using the majority outcome to cut single-run noise. Treat that as a practical floor and not a statistical guarantee. When a comparison matters a lot, collect repeated observations and show the counts, or call the difference directional. Keep the run dates and versions so you can audit later comparisons.
Be careful about cause, too. If you updated a product page and the next window looks better, that's worth noting, and it still doesn't prove the update caused it.
Done when: the report states its collection coverage and known limits, and no apparent change is described as proven cause.
What to do next
Pick three to five products and set up the registry, the prompt mapping, and one stable prompt set for each. Run a full cycle on the engines your plan covers, and use the first report mostly to find gaps in the registry and the labels. The second cycle is when the change columns start to mean something, and by the third you'll have a baseline you can defend.
If you're setting up the structure for several brands as well as several products, who owns which decisions and approvals is a separate question, and a governance model for multi-brand AEO programs covers it. This guide stays on the tracking structure and the reporting.
To see how the scheduled collection, answer history, and exports work with your own prompts, you can start a free DeepSmith trial and load your first prompt set.
Related guides
- What Your CMS Needs to Support for AEO: Evaluate whether your CMS supports AEO by checking structured data and schema control, clean semantic markup, fast updates, redirect and metadata handling, and what to ask vendors before.
- How to Do Keyword Research for AEO: From Keywords to Questions: Turn an existing keyword list into the questions people ask AI engines by expanding intent, phrasing conversational variants and grouping them by topic, so you can plan content.
- How to Monitor LLM Brand Mentions Without Buying Software: Monitor brand mentions across LLMs using a spreadsheet and a fixed routine.
- Personalization and Location in AI Answers: Why Two Users See Different Results: Learn how account history, region and language shift what an AI engine says, and what that implies for how far one tracked result generalizes.
- How to Set an AEO Budget for an Early-Stage Startup: Set an AEO budget for an early-stage startup by splitting spend between tracking, content and technical fixes, deciding what to defer, and tying each line to a milestone.


