AI will describe your product correctly and still recommend it to the wrong shopper

Picture a parent asking an assistant for a gentle laundry detergent for a newborn's clothes. The answer names your brand. It says, correctly, that you make a plant-based detergent. Then it says the product is "great for heavy-duty stains and workwear," because that is how a few retailer listings and a comparison site described it. The parent moves on to the next name.

Your visibility dashboard records a win. You were mentioned. The shopper who mattered heard the wrong thing.

That gap, between being named and being described right, is the subject of a study published on November 5, 2026 in Search Engine Land by Emma Baker. It is about dental practice software, not groceries, and we will be careful about what carries over. But its central finding is the clearest evidence we have seen for an argument we have been making for a while: counting mentions measures the easy part.

What the study measured

The researchers wrote 37 buyer-style prompts in five types (brand questions, category questions, segment fit, attributes, and head-to-head comparisons) and ran them across Google AI Overviews, Microsoft Copilot, Perplexity and ChatGPT. They tracked five software brands and scored each brand's treatment in each answer, about 2,000 scored mentions in all.

Each mention was scored on five parts of a positioning statement: category, ideal customer, value proposition, differentiation and proof. Before running anything they wrote an answer key from each brand's own published materials, so "correct" meant "matches what the brand says about itself."

The order in which positioning falls apart

The finding that matters most is that errors were not random. They followed what the authors call a fidelity gradient.

Part of positioning	What the study found
Category	Errors in fewer than 1% of full brand descriptions; average score rounds to 2.0 out of 2
Value proposition	Averaged 1.78 out of 2, helped by broad promises shared across the category
Differentiation	Varied by brand: correct in 87% of mentions for one, 43% for another
Ideal customer	Identified correctly in 60% of scored responses
Proof	No proof in half of full descriptions; full credit in 14%

Source: Search Engine Land, November 5, 2026.

Read top to bottom, that table is a list of the things that make a shopper choose, in reverse order of how well AI repeats them. What you are: nearly perfect. Who you are for and why anyone should believe you: shaky to absent.

Show the data

Measure	Share
Most-mentioned brand: named where it belonged	91%
Ideal customer identified correctly	60%
Outputs with at least one invented fact	34%
Full descriptions with full proof credit	14%

Source: Search Engine Land, November 5, 2026. Each bar is a different measure from the same study, not parts of one whole.

The most striking single example: Curve Dental appeared in 91% of the answers it belonged in, the best presence of any brand tracked. Yet only about a third of those answers got its ideal customer, its differentiation and its facts right together. By a mention count, it was winning. By what the answers said, it was mostly being described in someone else's words.

Why "who it is for" is the first thing to go

The study's explanation is worth quoting in spirit. When buyers asked which product suited a practice of a given size, the engines did not look up whom each brand said it served. They built their own sorting scheme and slotted brands into it. Brands that serve one narrow segment survived this well. Brands that serve several got squeezed into whatever slot was left.

Packaged goods have the same problem with different nouns. A sunscreen made for both kids and adults with sensitive skin. A protein bar positioned for both endurance athletes and people managing blood sugar. A detergent for babies' clothes that also handles gym gear. When an assistant builds its own grid (best for kids, best for sport, best budget pick) a brand with a layered position gets put in one box, often the wrong one, and disappears from the others.

We do not think the fix is to narrow your positioning to please a model. We think it is to state each audience plainly, on the page and in the places the model reads, so it has something to repeat other than a guess. That is the core of tip 3: one plain sentence saying what the product is, who it is for and what makes it different. The study suggests that sentence matters more than we had assumed.

Proof lives in other people's sentences

The proof finding should worry anyone who spent money on a homepage stats band. A software brand's homepage claim of more than 80,000 users appeared in 3 of the 384 answers that mentioned the product. A line from a press release, by contrast, was repeated across multiple answers, at one point inflated beyond what it said.

The authors' summary is blunt: proof in press releases and third-party coverage got picked up; proof in homepage stat blocks did not. That fits a separate analysis Search Engine Land published a day later, which argued that answer engines discount self-promotional claims unless independent sources say something similar.

For a CPG brand, the equivalents are easy to name: the certification, the award, the clinical test, the dermatologist association, the "number one in the category at a given retailer" claim. If it lives only in a badge graphic on your homepage, the study suggests most answers will not mention it. If it lives in a plain sentence on a certifier's site, a retailer listing and a trade article, it has a chance. Tip 21 covers how to show certifications so they can be read.

A gap does not produce silence. It produces a guess

The finding we would put on a sticky note is this one: 34% of scored outputs contained at least one invented fact. Price was the biggest source. One brand that does not publish pricing was quoted at least six different monthly prices across engines. In total, 223 mentions carried a price the researchers could not verify, and Copilot wrote 153 of them. Only 18% of the responses about that brand said its pricing was quote-based, which was the true answer.

Packaged goods brands do publish prices, but they leave other gaps that matter as much to a shopper. Which sizes exist. Whether the formula changed. Whether the product is fragrance-free or just lightly scented. Whether it is safe for septic systems, or contains a common allergen, or is made in a shared facility. When those answers are not stated plainly anywhere an assistant reads, the study suggests you should expect one to be written for you.

The study also tested a partial fix. A brand that publishes clear prices had its prices quoted accurately. Another that also publishes prices was still quoted in several different versions, and at least one traced back to a third-party directory rather than the brand. Publishing the fact helps. Keeping it consistent everywhere it is copied helps more. That is tip 5, and tip 25 covers what to do when an assistant gets it wrong anyway.

Each engine fails in its own way

The study gave each engine a "failure signature." Google AI Overviews compressed brands into one-line positioning and usually dropped the ideal customer. Perplexity was the most factually accurate but merged several brands into one shared clause. Copilot invented the most, including prices and unsupported negatives, with an unverifiable piece of information in 71% of its responses. ChatGPT gave the most complete descriptions, which also made its errors the easiest to believe.

We would not lean on those exact numbers for a grocery category; the sample was software, and engines change monthly. The general point holds: one answer from one assistant is not a measure of how you are described. The study found that a brand's presence changed between runs of the same prompt a third of the time.

What we think CPG teams should measure instead

Here is our position. Mention rate is a necessary number and a misleading headline. A brand that shows up everywhere and is described as being for the wrong shopper is not winning. We would add three measures to any AI visibility report, and we would weight them above the mention count.

Audience fidelity. For each answer that names the product, does it say who the product is for, and is that who you say it is for? Score it yes, partly or no.

Fact fidelity. Does the answer state any size, price, ingredient, claim or availability that is wrong or that you never published? Count the errors and note where they seem to come from.

Proof carried. Does the answer give any reason to believe you, such as a certification, an award or independent testing? If it never does, your proof is not in a form the model reads.

A practical version needs 20 to 30 shopper questions, three or four assistants, an answer key written from your own pages, and a repeat every few weeks. Some teams do it in a spreadsheet; XENA Intelligence runs shopper-style questions across assistants and scores how products show up, which is the same idea at a larger scale. The tool matters less than the answer key. Without one, you are counting mentions again.

We argued in Stop asking where you rank in ChatGPT that a single rank is the wrong unit. This study adds a sharper version: even a perfect presence score can hide an answer that sends your best customer somewhere else. And as our piece on agents repeating listing claims noted, the words an assistant uses about your product increasingly become the sales pitch itself.

The limits, stated plainly

This was one B2B software category, five brands, four engines and one research team's scoring rubric. The authors did not test fixes, and they say so. Packaged goods differ: shoppers ask shorter questions, retailer listings carry more weight, and reviews say more about who a product suits. We would not quote 60% or 34% as facts about your category.

What we would take is the shape. Category is safe. Audience and proof are fragile. Silence gets filled. If that shape holds for a product as easy to describe as practice software, we see no reason to expect it to be kinder to a detergent with three audiences and a reformulation last spring.

The study's closing line is the right instruction for a brand team: AI engines will describe your brand whether or not you give them the material to do it well. The work is making sure the version they repeat is yours. A tool such as SearchPanel, a spreadsheet or a weekly habit of asking the questions yourself all count, as long as someone is reading what the answers say and not just whether you were in them.