what happens when someone asks ai about your brand

Ask AI to recommend within a category and the language it uses about the brands it names is close to uniformly positive. Across 110 brands, 88.0% of the perception themes AI expressed in category search were positive, and 0.6% were negative. Ask the same models about one of those same brands by name and positive themes fall to 72.3% while negative themes rise to 5.1%, close to nine times the category-search rate. Mixed themes, where the model both praises and challenges the same trait, rise from 0.9% to 8.4%.

This is not a few brands dragging an average. 81% of the 110 brands show the effect individually, and it holds whether themes are counted equally or weighted by how often they came up.

A note on what that percentage counts, because it is easy to overstate. AI does not hand back a sentiment score. It returns sentences, and Centium summarizes the roughly 30,000 sentences that come after a brand is mentioned into about 13 recurring perception themes per brand, each carrying a sentiment and a frequency rating. The percentages above are shares of those themes across the Centium portfolio, weighted by frequency. Every sentence feeds into a theme, with the data reported from the consolidated themes.

There is also a structural reason the two numbers differ, and it is worth stating plainly rather than burying in a method note. In category search a brand is only described if the model already decided to recommend it, so the description is filtered by that decision. Brand search applies no such filter. That is the finding rather than a flaw in it: the sentiment a brand sees in category search is what the model says about a winner it has already picked, and the sentiment in brand search is what a buyer gets when they look the brand up cold. Those are two audiences seeing two different accounts of the same brand.

The size of the gap is one thing. What fills it is another. 96 of the 110 brands, 87%, carry at least one negative or mixed theme in brand search, and the concerns cluster tightly.

Price is the landmine, and it is not close. More than half the brands, 56 of 110, have AI raising cost when someone asks about them by name. Most of that is not an outright complaint but a hedge, the model praising the product and then qualifying it on price. Quality consistency follows at 27% of brands and customer service at 22%.

Two of these barely exist in category search. Customer service and fulfillment comes up for 22% of brands in brand search and for no brands at all in category search. Themes about the company and its workplace, covering employee culture, ownership changes, and general reputation, run at 11% and again at zero. When AI is choosing between products it never raises how a company treats its staff or answers its phone. Ask about that company by name and both surface.

That is the cleanest statement of what changes between the two modes. Category search criticizes the product, with reservations about weight, noise, price, and what is in stock. Brand search criticizes the company.

Before it answers, a model runs its own searches, a behavior called query fan-out. Report 01 measured 322,485 of them to establish what AI looks for. Here the same classifier is pointed at the mode split: 157,476 fan-out searches from category prompts and 15,405 from brand prompts, across 111 brands.

The largest single change is what stops. In category search, 26.5% of the searches AI runs contain the word best or top, and 87.3% of prompts trigger at least one of them. In brand search that collapses to 2.2% of searches and 11.1% of prompts. All 111 brands show the drop, without exception. The model is no longer shopping.

What takes its place is scrutiny. Searches that hunt for a problem, the ones containing complaints, issues, scam, legit, worth it, cons, drawbacks, lawsuit, or recall, run at 0.3% of category searches and 3.7% of brand searches. On the prompt measure the gap is wider still.

Report 01 flagged this behavior on branded search and called it the clearest sign that AI vets a reputation the way a skeptical buyer would. Separating the two modes shows how cleanly it splits. Hunting for the downside is almost purely a brand-search move, and it is the single behavior that best explains everything in the two sections above.

Three smaller shifts point the same way. The model reaches for named communities and platforms about twice as often, and on the prompt measure that runs from 15.3% to 29.7%. It runs head-to-head comparisons about twice as often, and searches for alternatives roughly six times as often, though from a very small base. Comparison, alternatives, and community discussion are all ways of asking the same question: is this one actually any good, and compared to what.

One intent does not move at all. Review-seeking runs at 20.0% of category searches and 19.8% of brand searches. Whether it is picking from a field or checking a single name, the model wants to hear from someone who has used the product. Review coverage is the one asset that pays in both modes.

Two other things change. Recency drops hard: naming a recent year or asking for the latest all but disappears once a brand is named. When the model is choosing between options it wants this year’s answer; when it is checking a company it is reading a record that does not expire, and an old complaint counts as much as a new one. Brand search is also narrower, running 8.8 searches per prompt against 14.2 in category search, so it covers less ground with each search aimed more precisely.

The sources follow the search. Reading only the citations AI returned from a live web search, across 113 brands, the mix moves in three ways once a company is named.

First, the brand’s own site comes up. Centium reports sourcing as a usage rate, the share of responses citing a given site, and on that measure a brand’s own site is cited in 11.2% of category-search responses and 69.3% of brand-search responses, and 112 of the 113 brands show the increase. When the model is comparing options it mostly reads other people. When it is checking one company it first goes to the source. In category search AI does read brand-owned pages, they just belong to other brands.

Second, the field narrows. Category search draws on about 105 distinct domains per brand against 76 in brand search, and the top ten domains carry 37.9% of category citations against 49.9% of brand citations, so it is drawing on fewer sources and giving each of them more weight.

Third, the kind of publication changes.

The names that arrive are the same ones every time. LinkedIn leads at 7.5% of brand-search responses, and Trustpilot climbs from 0.1% in category search to 6.0%. Behind them sit Yelp, TripAdvisor, and the Better Business Bureau, then the company-record sources category search barely touches at all: ZoomInfo, Indeed, and PitchBook. Those are review aggregators, corporate records, and employer profiles, the sources you would pull to vet a company rather than to buy from one.

Underneath them sits a backbone that does not move. Reddit and YouTube are each cited in a fifth to a quarter of responses whichever way the question is asked, dwarfing every source in the two groups above. Wikipedia sits further back but more than doubles once a brand is named, which is its own small piece of the same story. Social and community sources rise in aggregate too, from 10.7% of category citations to 14.4% of brand citations, yet the platforms themselves show up heavily either way. That makes them the most reliable surfaces in AI visibility and the least mode-dependent: everything else about how AI searches changes, and these stay.

None of the sources in the section above belong to a particular industry. Your own site, LinkedIn, Trustpilot, Reddit: the same short list is in play whatever you sell. Category search works the other way, and to see it you have to stop averaging across the whole portfolio. A cycling title never appears for a skincare brand, so measured across all 113 it looks small. Measured against the brands it actually covers, it is one of the largest sources in this report.

The recurring names are the ones that test and rank things, and they change completely from one industry to the next. BikeRadar reaches 32.2% of category-search responses for cycling brands, and the pattern repeats straight down the list: outdoor gear has OutdoorGearLab, beauty has Allure, software and B2B has Gartner, food and packaged goods has Healthline, travel has Condé Nast Traveler. Each runs near the top of its own industry and is close to invisible outside it.

That is the sharpest structural difference between the two modes. The brand-search sources barely vary by sector, so a cycling brand and a software company get pulled toward the same kinds of places: their own site, the professional networks, the review aggregators. Category search fragments into a different set of publications for every industry, with almost no overlap between them. One target list travels from industry to industry. The other has to be built for each.

Put the findings together and the mechanism runs end to end. Ask about a company by name and the model goes looking for a problem. It lands on the sites where problems are recorded, the review aggregators, the complaint registries, the employer profiles. Then it comes back measurably less positive about the same brand it had praised in category search. The searches, the sources, and the sentiment are one behavior measured three ways.