How to tell whether an AI engine is comparing your product against the right competitor set
On this page
A premium espresso machine, the kind built for a serious home setup, gets asked about in ChatGPT. The answer comes back as a table, and the two other rows in it are entry-level drip machines a third of the price. The comparison isn’t wrong because the engine picked a worse product. It’s wrong because it picked the wrong shelf to compare from. Nobody designed that outcome. It fell out of how the table got built, and the fix is different depending on where in that process it went wrong.
The engine assembles the competitor set, you don’t
A comparison table in ChatGPT, Perplexity or Gemini isn’t a human editor’s shortlist. It’s generated per query from whatever products the engine can confidently match to the question and the attributes it can state something specific about. Evertune’s review of ChatGPT’s new table layout, introduced on March 24, 2026 and already showing up in close to a third of roughly 21,000 shopping-style responses it sampled days later, found the rows are category-specific: Wi-Fi standard and setup ease for routers, noise cancellation and battery life for headphones, bag size and ingredient quality for pet food (Evertune). The table is built from whichever products carry that category’s attribute set in a form the engine can read, not from a curated sense of who your real competitors are. If your product’s category or tier isn’t stated plainly somewhere the engine reads, it gets slotted in with whatever products do state it clearly, which is not always the tier you actually compete in.
Our piece on what changed when ChatGPT, Perplexity and Gemini became shopping engines covers the table format itself in more detail. This is about what happens once the table exists: specifically, whether the peer group it assembled is the one you’d pick.
Comparisons are the least reliable thing these engines do
This isn’t a minor edge case. Product.ai ran 220 shopping questions, covering laptops, TVs, mattresses, sunscreen and robot vacuums, against the free and paid tiers of ChatGPT, Claude, Gemini and Perplexity, five times each, for 8,794 responses. Across the whole set, 86% of questions produced a repeatable factual conflict. But the error rate wasn’t even across question types: straightforward factual or spec questions produced a conflict 75% of the time, while questions asking the model to compare two products head to head produced one 97% of the time (Product.ai study, reported by dnyuz.com). Comparisons are harder for these models than single-product facts, by a wide margin, and a wrong competitor set is one concrete way that extra unreliability shows up: not just a wrong number in the table, but the wrong products sitting in it to begin with.
Three places the wrong competitor gets in
Category and tier data that doesn’t say what it should. If your product page, feed and structured data describe the category (“espresso machine”) but not the tier within it (price band, build class, intended user), the engine has nothing to sort you by except whatever else is in that broad category with clearer signals. A $2,000 machine and a $70 one both satisfy “espresso machine.” Only one of them states, in a form the engine can read, that it belongs on a different shelf.
Variant and identity confusion. Smarter Sorting ran 2,500 interaction steps across five assistants and ten product categories and found 71.8% of steps failed to achieve full product-truth reliability, with variant confusion, the engine conflating a specific SKU with a different one from the same line or a different brand entirely, showing up 234 times in that test (Smarter Sorting). A table built around the wrong variant of your own product is also, from a buyer’s view, a table comparing against the wrong thing: the version being evaluated isn’t the one you sell.
A query that doesn’t specify the tier either. “Best espresso machine” doesn’t say premium or budget, so the engine has to infer which shelf the asker means, usually from whichever products have the most unambiguous attribute data for that inferred shelf. This is the one source that isn’t really a data problem on your end. It means the same product can get a correct competitor set for one phrasing of a question and a wrong one for a near-identical phrasing, which is itself diagnostic: if the set flips with the wording, the cause is query ambiguity, not your product data.
Check which one is yours before you fix anything
Run the comparison query a handful of ways: with your product’s actual tier stated (“best premium espresso machine”), without it, and with a direct ask (“compare [your product] to its closest competitors”). If the competitor set stays wrong even when the tier is stated outright, the problem is upstream of the query, in how your category and tier are expressed on the page and in your schema; building the entity signals AI engines trust covers making that kind of structured fact legible rather than implied. If the set corrects itself once the tier is named, the issue is query ambiguity rather than your data, and the practical fix is making sure your own content answers the tier-qualified version of the question, since that’s the version that gets you compared correctly.
Tracking competitors in AI answers is how you see the actual competitor set an engine is naming for you across more than one prompt, rather than relying on a single spot check, and watch whether it’s consistent or shifts with phrasing. If the wrong product in the table is one of your own SKUs rather than a rival’s, the mechanism is adjacent but the diagnosis is different; why an AI shopping assistant picks one of your products over another covers that case.
Don’t guess which shelf you’re being measured from
A comparison table that seats you against the wrong peers doesn’t just make you look worse on a few attribute rows, it measures you against a standard you were never trying to meet. Run a free Visibility Check on the comparison query where you suspect this is happening, and look at who else is in the table before you look at how you scored against them. If the wrong set shows up more than once, Monitor tracking will tell you whether it’s a one-off phrasing problem or a standing gap in how your tier is stated, which is the difference between leaving it alone and rewriting a product page.