For about twenty years, brand discovery worked one way. Someone typed a question, got a page of links and chose. Every piece of digital marketing infrastructure was built around that choice.
That is no longer how a growing share of category questions get answered. Someone asks an assistant and gets one answer: a paragraph, perhaps three brands, and a recommendation. Comscore describes the shift as a move from search behaviour to answer behaviour.
We put this to a panel at our recent webinar: ShiSh Shridhar, Global Director for Retail and CPG Startups at Microsoft for Startups, with i-Genie.ai’s Head of Data Science Linda Hoeberigs and Head of Product Agne Kasulaite.
Nobody chose those brands. A system did.
Ranking first on Google used to be the prize. Ahrefs found that when an AI Overview appears on an informational search, the top result’s click-through rate falls by around 34.5 per cent. And SparkToro reports that two in three US Google searches in early 2026 ended without a single click.
That changes which brands enter consideration, how yours is described, and whether you notice before it shows up in share.
Why one AI answer tells you almost nothing
Ask an assistant the same question twice and the answer moves. That is a property of how these systems work, not a fault. Here’s what Linda had to say on the topic during our recent webinar. To watch the full discussion, click here.
A 2026 study of 12,933 AI responses across 20 brands, 8 languages and 3 models found that query language explained 26.5 per cent of the variation in a single answer, against 1.5 per cent for brand. The language your customer asks in moves the answer nearly 18 times more than the brand they ask about. The reliability of a brand ranking drawn from one answer was around 0.01, on a scale where 1 is perfect.
“Your brand will have a different reputation in every language you sell in. If your brand team is only checking in English, you don’t know how you’re being described in the markets you can’t read.”
Linda Hoeberigs, Head of Data Science, i-Genie.ai
So one screenshot of an assistant naming your brand tells you nothing, and five tell you barely more. The fix is sampling design, not better prompting. Ask any AI visibility vendor about their sampling before you look at their dashboard: which assistants, how many questions, which prompts and what error margin. ShiSh counts around 50 startups selling this, and a single answer is a coin flip. That is why i-Genie.ai reads millions of AI answers every day across six assistants.
Measuring brand equity when AI does the describing
AI answers come either from training data, frozen at a cut-off date, or from live sources the model retrieves and synthesises, which can shift within days.
Which route you get depends on the question. Ask what something is and the assistant answers from memory. Ask which one to buy and it goes and looks it up. So the answers that matter most commercially are the ones that move fastest.
Those live sources are mostly not yours. A study of AI citations about brands found that 85.7 per cent pointed to sites the brand does not own. Reviews, retailer pages, editorial and community discussion do most of the talking, and that third-party consensus carries far more weight than anything on the brand’s own website.
“The uncomfortable truth about LLMs is that they’re not forming an opinion about your brand. What they’re doing is reading your reviews out loud.”
Agne Kasulaite, Head of Product, i-Genie.ai
So measure what the answers are made of: search behaviour, social conversation, reviews and video. These signals sit upstream of the answer, they are far more stable, and when they move you can see why.
Measure the whole layer, not one channel. In vitamins, i-Genie.ai found the same brand leading AI visibility in both Australia and India while weak on search in both. Judge it on search alone and you would call the wrong winner.
That does not make the models useless. In Linda’s words, a model is the reader of the information, not an oracle. It can read millions of reviews and posts across markets and languages, and say why a review is negative, not just that it is. But it is only as trustworthy as the data it is pointed at, which is why i-Genie.ai built Presto on its own data rather than on whatever a general model absorbed.
Brand tracking metrics for AI answers: the four Ps
Four things can be tracked honestly, and they are not equally useful.
-
- Presence. Does the brand appear at all? Easiest to measure, least informative
-
- Prominence. How often and how high you appear against competitors. Also easy to measure, and close to a vanity metric: being named first is not the same as being recommended
-
- Perception. Which attributes get attached to your brand, and how that compares with competitors. This is the equity signal
-
- Provenance. Which sources keep turning up behind the answers. The one you can act on directly
Perception is where the real finding sits. Being named as the affordable option is a very different position from being named as the one dermatologists recommend, yet both count as presence.
Assistants rarely bash brands, but they describe them very differently. Ask a million haircare questions and the same brands keep appearing. The surprise is that a challenger can be described more favourably than the market leader, because it did a better job of curating reviews on retailer sites, giving the assistant cleaner claims to pull.
Provenance can cut the other way too. In the menopause category, i-Genie.ai found a client’s own content being cited in answers about a competitor’s brand. Spotted early, that is fixable, for example by naming your brand more prominently in that content.
When the agent becomes the shopper
Retailers such as Walmart now run their own shopping assistants, answering from their catalogue, ratings and reviews. That turns your product page from a listing into source material.
On 21 October, Walmart opened an Agent Gateway, an MCP server that lets approved AI agents search its inventory and buy on a shopper’s behalf, with Agent Profiles and spending limits following in November. Amazon, meanwhile, is blocking third-party shopping agents. Two of the biggest retailers have bet on opposite futures.
Either way, an agent does not get persuaded. It reads attributes: ingredients, pack sizes, claims, price, stock and ratings. When the agent fills the basket, complete and consistent product data does the job the shelf used to do.
“Product pages are designed by the e-commerce group for a human user to go in and convert to a sale. They’re not designed for an agent.”
ShiSh Shridhar, Global Director for Retail and CPG Startups, Microsoft for Startups
Consistency matters just as much as completeness. As ShiSh highlighted, when retailers provide conflicting information about the same product, AI agents may look elsewhere for more reliable sources. Here’s what Linda, Agne and ShiSh had to say about this during our recent webinar. To watch the full discussion, click here.
Four brand tracking steps before your next planning cycle
-
- Baseline before you build. Establish a baseline in the signal layer, not from a handful of prompts
-
- Perception over presence. Measure how you are described, not just whether you appear
-
- Provenance before prompts. Trace the sources behind the answers and fix the ones you own or influence first, starting with retailer product pages, then recent reviews, then content AI can read easily. Give each an owner. Tactics that try to manipulate what a model says carry reputational risk, especially in regulated categories. If an assistant says something non-compliant, it is probably repeating a source: trace it and fix it there
-
- Proof before praise. Re-measure before you claim an intervention worked
This is the layer i-Genie.ai was built to read. Brand Pulse tracks equity continuously from search, social, reviews and video, against every competitor, with two years of back data from day one. For the full argument, read our white paper on augmenting brand equity tracking with digital signals, or watch the replay.
