The screenshot arrives on a Monday morning. The CEO was showing a friend what ChatGPT says about the company, and the answer describes your brand with the positioning you retired three years ago. Or it recommends your competitor as the obvious choice and does not mention you at all. The note under the screenshot is one line. “Is this right?”
You have an opinion about it. You do not have a number.
Your brand tracker has a page for every audience you have ever cared about: customers, prospects, the trade, the press, your own employees. It does not have a page for the assistant that just answered the CEO. You can prove the repositioning landed with people; the tracker has three years of it. You cannot prove it landed with the assistants those people now ask for advice. Nobody in the building can say what they believe about the brand, whether the belief is getting better or worse, or whether the free tier and the paid tier of the same assistant even agree with each other. They often do not.
This is not a channel problem, and the tools rushing in to treat it as one are counting the wrong thing. AI assistants are an audience. They have formed an opinion of your brand. And they are the only audience your brand has never put a number on. By the end of this you will have a list you can run yourself, with or without us.
The customer moved first
Scale is what turns this from a digital-team question into a board question.
OpenAI said in February that ChatGPT has 900 million weekly active users. Google’s AI Overviews, the summary that now sits above the results on a search page, had already passed two billion monthly users by the middle of last year. Bain found in early 2025 that about 80% of search users already relied on AI summaries for at least 40% of their searches on traditional search engines.
The volume is half of it. The other half is the shape of the answer. The sales trainer John Barrows put it in one line on his podcast this summer: “we’ve gone from this search engine to an answer engine, and that’s a dangerous thing.” He meant the danger to the person who stops asking follow-up questions. There is a second edge, on your side of the answer. A results page offered ten links and let the customer choose. An answer offers one recommendation and a reason. If the reason is wrong, it is wrong for everyone who asks.
For the commercial side of the house this is a demand question: which brand does the assistant put first, and why. For the brand side it is a perception question: what does the assistant believe about us, and where did it learn it. Underneath, it is the same question, and nobody has the instrument.
Where the opinion comes from
The mechanics matter, because they decide what you can fix and how fast.
An assistant answers from two places. The first is what it absorbed in training: years of the open web, reviews, press, forums, your own site as it was, your competitors’ sites as they were. Call that memory. The second is what it finds when it searches at the moment of the question. Call that the live read. The free tier of an assistant and the paid tier of the same assistant do not always say the same thing about you, and the assistants do not agree with each other.
David Friedberg made the point on All-In in August, about how companies will buy AI models: “the idea that there’s kind of a model that you pick for everything, I think is the false assumption.” He was talking about enterprise procurement. For a brand it means something simpler. There is no single assistant to check. Your customers are spread across several, on free and paid tiers, and the assistants hold different opinions of you.
The two sources are two different problems. A gap in memory is a reputation problem. It was written into the model over years of what the web said about you, and content you publish tomorrow will not move it for a long time. A gap in the live read is a content problem. It is about what the assistant finds when it looks, and that can be changed this quarter. They do not share a fix. Any single score that blends the two hides the one thing you needed to know: which kind of work is ahead of you.
This is also why one person on the team checking one assistant once is not measurement. It is an anecdote with a screenshot, and the company will treat it as the truth until the next screenshot.
The industry’s answer is to count
The market has noticed, and a category of software has grown up around it almost overnight. Every one I have looked at sells a dashboard. Mentions. Share of voice. A visibility score. Some citation counts. It’s the first generation of social listening all over again: count the volume, watch the line, call it insight.
A mention count cannot tell you what this audience believes. Being named in most answers says nothing about whether the answer called you the premium choice or the budget option, whether it associated you with the thing you spent a decade building, or whether it described a company you stopped being years ago. You would never accept a share-of-conversation number as a brand-health read for any human audience. It is being sold as one for this audience because it is the easiest thing to count.
And the count itself is becoming free. Spenser Skates, the CEO of Amplitude, built an AI visibility tool, gave it away, and said why on the Lightcone podcast last December: “I think the real business has to be downstream of AI visibility… it’s so easy to do… And so the commoditization is going to happen real, real fast.” He’s right. The scarce thing was never the number. It is knowing what the number means for your positioning, and what to do about it.
Meanwhile the people who own brands mostly believe this is handled. In a survey of 602 marketing and PR professionals published in July by Scrunch and Scribewise, two firms that sell into this market, 91% said their team has a clear, documented strategy for AI search visibility. In the same survey, 70% were not monitoring sentiment toward their brand in AI answers, and 71% were not analyzing AI share of voice against competitors. The report’s own title says it: moving fast, flying blind. Strategy exists. Measurement does not.
Treat it like an audience
Brand strategy has always rested on how the audiences that matter perceive you. You do not measure customers by counting how often they say your name. You ask what comes to mind first, what they associate you with and how warmly, whether you stand apart from the set they compare you to, whether you feel right for the need, and who you are to them, facet by facet. Every brand leader reading this has run that battery for years, against a positioning they wrote down on purpose.
Now the strange part. This new audience answers in full sentences. It never gets survey fatigue. It never tells the interviewer what it thinks she wants to hear, because it does not know it is being interviewed. It never forgets where it learned something, and it will show you its sources if you ask. Put the questions your customers actually ask to it and it will tell you, verbatim, what it believes about your brand and what it read to believe it. It is the most legible audience you have ever had.
Nobody knows what it thinks because nobody has asked it properly.
Asking properly looks like this. Real customer questions, scores of them, the ones people in your category actually type, not five prompts someone tried on a Friday. Both reads, memory and live, kept separate, across the assistants your customers use, on the tiers they use. The answers scored on the measures you already run: salience, associations and their tone, differentiation and relevance, identity. Where a measure cannot be read in this channel, say so. Loyalty is one. An assistant cannot tell you whether a customer will come back, and a scorecard that guesses is worse than one that admits it.
Then the part that most measurement skips: a yardstick. Jeff Dean, Google’s chief scientist, said something in August about instructing AI agents that holds just as well here: “the importance of specifying what it is you want has actually gone up.” You cannot measure drift from a positioning you never wrote down for this audience. The yardstick has to come from your own people: what you intend to stand for, the products that matter most, the customers you want to win, the brands you count as your set. The answers are already in your company. The assistant’s answers get measured against them.
Two more things, and they are the two that separate a measurement from a mood. Measure twice, independently. The assistants do not answer identically twice, so a single run gives you a point with no sense of its own spread. Two independent runs give you a range, and next quarter’s change has to clear that range before anyone is allowed to call it movement. And state the boundary: what an assistant says is evidence of what its training data and the retrievable web say about your brand, not what consumers themselves think. Keep the human tracker running. Read the two side by side. The gap between them is a finding in itself.
How Calafai approaches it
I will describe what we built, so you can judge the approach rather than take my word for the argument.
We built two instruments and we run them in an order. Calafai Brand Standing answers the strategy question: what does AI believe about your brand, measured against what you intend it to stand for, on the measures brand strategy already runs on, with the evidence attached to every score and an error band from two independent measurements. It starts with a working session with your team, because that is where the yardstick comes from. It ends with a scorecard, the drift between your positioning and the one the assistants repeat, and a list of what to start standing for, what to stop claiming, and what to keep defending. A senior brand strategist reads and signs off on every report before it reaches you.
The Calafai Brand GEO Report answers the operational question that follows: where do the assistants still tell a different story, question by question, which sources did they read to say it, and what do you fix first. (GEO is generative engine optimization, the work of changing how generative AI describes you.) It is a plan rather than a dashboard. Each gap comes with the page, the listing, or the source that needs to change, with the live-read gaps first because those are the ones you can move this quarter. Most brands end up needing both, in that order: decide what AI should say about you, then fix where it still says something else. Where the assistants still tell a different story, and what it costs you when they do, is the companion piece, The Conversation Before the Click, which publishes on Thursday, September 17.
Both rest on the same method. Real questions, put to the real assistants, every answer kept, the two reads never averaged. Nothing is modeled or estimated, and the Brand GEO Report names exactly which assistants ran for your brand. We built strategy software for our own work first, the Calafai platform, and we run these engagements on it, which is why a measurement this thorough takes weeks instead of a quarter. AI agents do the measuring at scale. The judgment stays with people, ours and yours, and you direct the work. Your positioning brief never trains a model.
The habit behind all of this is older than Calafai. I spent a decade inside a global brand, building the analytics and intelligence tools its teams ran on, and before that I built analysis software for the U.S. intelligence community, where a finding that could not be traced back to its source did not count as a finding. A score without the answers behind it is an opinion with a decimal point.
What I would do this quarter, whoever you call
If you never call us, do this anyway.
- Write down what you want the assistants to say about you before you measure anything. Positioning, the set you belong to, the customers you want. One page.
- Keep your human tracker running next to whatever you measure, and read the two side by side.
- Fix the live-read gaps first, this quarter. Work the memory gaps through what gets written and cited about you, and expect that to take longer.
- Put a person’s name on the reading. A dashboard nobody signs is a dashboard nobody defends.
The part that compounds
None of this is urgent in the way a campaign is urgent. It is urgent in the way a reputation is. Every answer an assistant gives feeds the next one, because what gets written and cited about you is what the assistants absorb next. A description left alone hardens.
The next screenshot is coming. The note under it will say the same thing. “Is this right?”
You can have that answer before it arrives. This audience answers in full sentences, names what it read, and never asks you to take its word for it. It has been answering all along. Nobody has been asking.
If you want to argue with any of this, or find out what the assistants say in your category, write to [email protected].
Sources
- Aisha Malik, ChatGPT reaches 900M weekly active users, TechCrunch, February 27, 2026. The 900 million weekly active users figure, from OpenAI’s announcement of the same day.
- Sundar Pichai, Q2 earnings call: CEO’s remarks, Google, July 23, 2025. “AI Overviews now has over 2 billion monthly users.”
- Bain & Company, Consumer reliance on AI search results signals new era of marketing, press release, February 19, 2025. “About 80% of search users rely on AI summaries at least 40% of the time on traditional search engines.”
- Scrunch and Scribewise, 2026 AI search survey: Moving fast, flying blind, published July 16, 2026. 602 U.S. marketing and PR professionals surveyed May 19 to June 2, 2026. The 91%, 70%, and 71% figures.
- John Barrows, Make It Happen Mondays, “AI Agents, False Productivity, and the Sales Team Reset with Gabe Larsen,” July 6, 2026. The “answer engine” quote.
- David Friedberg, All-In, “Google’s AI Brain Drain, SpaceX’s Huge Quarter, Airtable’s 90% Collapse, US Data Fuels China AI,” August 8, 2026. The “false assumption” quote.
- Spenser Skates, Lightcone Podcast (Y Combinator), “What Founders Have to Unlearn to Become Great CEOs,” December 3, 2025. The “downstream of AI visibility” quote, and Amplitude’s own account of building the tool and giving it away.
- Jeff Dean, Y Combinator Startup Podcast, “Jeff Dean: The 1% Rule for Building in AI,” August 1, 2026. The “specifying what it is you want” quote.
Podcast quotations are taken from transcripts of the recordings and lightly trimmed with ellipses; nothing inside a quotation has been altered.