What is an AI engine saying about your brand right now?
Its own version of the brand, pieced together from sources the brand did not choose and phrased in words the brand did not write.
Picture the head of marketing at a CRM built for mid-sized sales teams. One evening she types her buyers' question: "What's the best CRM for a 50-person sales team?" The engine names five products. Hers is fourth, described as the product it was two years ago, before the move upmarket: "a good budget option for small teams". The phrase comes almost word for word from a competitor's comparison page.
Nobody asked her. The story was told without her in the room, and tomorrow it will be told again, a little differently.
Where does that story come from?
From two places at once: the model's memory, fixed when it was trained, and the pages it reads when it searches the web today.
Memory is where old positioning lingers. A model knows a product through what was written before its training ended. A pivot or a new price after that point may not be there at all.
Search brings in today's web. Engines like ChatGPT, Claude and Gemini can search, and Perplexity always does. They read reviews, comparison pages, "best of" lists and forum threads, often written by someone other than the brand.
The two sources often disagree. In a Visibility Labs study reported in June 2026, ChatGPT's product recommendations changed by about 80% between web search on and off. Which view a buyer hears depends on whether the engine searched.
The engine also picks the criteria. A buyer rarely says what matters, so the engine decides: price per seat, say, or how the product copes with a team of fifty. A brand strong on criteria the engine did not pick gets little credit for it.
Why does the story change with the question?
Because the question sets the criteria, and the criteria decide the story.
Ask "What's the best CRM for an enterprise sales team?" and the engine weighs security and admin controls. Ask "What's an affordable CRM for a five-person startup?" and it weighs price and setup time. The same product can be the safe pick in the first answer and the overpriced one in the second.
Even one question does not hold still. Ask it again and the list changes; What is GEO? has the numbers.
A brand has thousands of small stories, not one. So a single check proves nothing either way. What counts is the pattern across many questions: the mention rate, meaning the share of answers that name the brand, and what those answers say.
How does the story go wrong?
In four recurring ways, with a plainer problem underneath them: some facts are simply wrong.
Searchable, a vendor that sells tools for this problem, graded about 1 in 10 brand facts in AI answers as wrong in its July 2026 report, pricing most often. Weigh the figure accordingly. The four failures below are harder to spot, because an answer can be accurate sentence by sentence and still tell the wrong story.
The mix-up
Two brands that look alike get blended into one, and a strength that belongs to one is credited to the other. We think the cause is how a model learns: from text, where two products described in the same words for the same buyers are hard to tell apart. We know of no study that measures how often it happens, so treat this one as a thing to check for, not a statistic.
Borrowed framing
A competitor's comparison page becomes the engine's description of you. Comparison pages are built around the buyer's own question, which is the shape engines like to cite: in an AirOps study from April 2026, ChatGPT cited headings matching the question 41% of the time, against 29% for the rest. When the only page comparing two products was written by one of them, the engine has little else to quote.
The most convincing description of your product may be one your competitor wrote.
The echo
Different engines may tell much the same story, so one weak version can travel everywhere. In a NeurIPS 2025 best-paper winner, researchers found that different AI models give strikingly similar answers to open-ended questions. It did not test brand questions, so read it as a warning, not a measurement: a second engine may be less of a second opinion than it looks.
The empty chair
The brand is simply not there. Asked for product recommendations with web search on, ChatGPT named about five products per answer in the Visibility Labs study. An answer has no page two. A buyer who reads five names has no reason to suspect a sixth.
What does leaving it alone cost?
The only hard evidence is about buyers, and it says they act on what the engines tell them. In G2's April 2026 survey, 69% of B2B software buyers had chosen a different vendor because of what a chatbot told them. G2 sells software reviews and has a stake in this story, so read the figure loosely.
A buyer who reads that a product is "the budget option for small teams" does not go looking for its enterprise features.
What can a brand do about it?
Read the stories at scale first, then trace each one to its source, then give the engine the brand's true story where it will find it.
Reading at scale means many buyer questions, each asked once of each engine, with every answer read for how it treats the brand, from a recommendation to a warning. Tracing follows each answer back to what the engine read and the criteria it judged by.
The aim is modest: the brand's true story, in the terms the engine uses to decide, in the places it reads. Memory is out of reach until the next model is trained. Search is not. The brand's own site comes first, because it is what the engine reads when it searches and the fastest thing a team can change. Then come the places others talk about the brand: reviews, comparisons, communities and video. Asking the same questions again shows whether the stories moved.
For the practical steps, read the marketer's guide to AI visibility. For how the loop runs, see How it works.
Sources
- G2, "Half of B2B software buyers now start their research with AI chatbots", April 2026 — source
- Visibility Labs via Search Engine Land, "80% of ChatGPT product recommendations change when search is enabled", June 2026 (20,000 responses) — source
- Searchable, AI brand misinformation report, 32,556 facts, July 2026 — source
- AirOps with Kevin Indig, "The fan-out effect", April 2026 — source
- "Artificial Hivemind", NeurIPS 2025 (best paper) — source