September 29, 2026
Before AI can recommend you, it has to know you and your brand. So what does it actually know?
Most conversations about AI visibility start in the wrong place. People jump straight to “how do we get cited?”
Fair question, but it’s the second question, not the first. The first one is simpler:
What LLMs know about a brand or product can vary wildly, depending on whether they’re answering from memory or looking you up on the spot.
That gap is where we typically start with clients.
When an LLM answers a question about your brand, it’s doing one of two things. Either it’s working from its training data, which is everything it soaked up from the open web before its cut off date, or it’s running a live search and reading a few fresh pages before it answers. The industry calls these training data and grounded responses.
How often each one happens is harder to figure out than it was a year ago, and it also depends on what you ask. Cloro looked at this in August 2026 and found ChatGPT searched the web for 86.5% of commercial prompts and less than 1% of informational ones. Graphite got 58% on prompts that compared companies, when ChatGPT was left to make its own mind up. Similarweb thinks only 7% of all US ChatGPT answers include a citation at all. Those numbers look like they disagree, but they’re measuring different kinds of questions. Ask “what is a B Corp” and the model answers from memory. Ask “which London restaurants are B Corps” and it goes and looks.
It varies by platform too; Perplexity searches all the time. Google’s AI Overviews and AI Mode are built on live search results rather than memory, though Overviews only show up on some queries.
Gemini seems to be the odd one out. It searches the least, and it’s inconsistent about it, anywhere from 27% to 70% of answers depending on the month. Claude searches when it decides it needs to, and on smaller brands it will often just say it doesn’t have the information rather than guess.
This matters because the two routes do different jobs.
Training data carries the long term story: who you are, where you came from, what you stand for.
Grounded search handles whatever’s current: prices, news, new openings, anything that moves. Being strong in one and weak in the other is a problem either way, just a different problem.
So when we audit a brand’s AI foundations as part of our SEO and AI search campaigns, we’re really running two tests side by side. What does each model already think it knows about you, and what does it find when it goes looking? Same questions, both modes, across the main platforms.
When it comes to optimising for AI search, we start with six diagnostic signals, each answering a different question. The value is running them together and seeing where the picture holds and where it falls apart.
Does the model correctly identify what kind of organisation you are? A restaurant group, an energy company, a membership body, a SaaS platform? This might sound like the basics but models can still get it wrong, particularly for brands whose name overlaps with something else.
Founder, founding year, founding city, ownership, scale. When these are right, everything downstream tends to hold. When they drift, you can usually trace a content or PR gap underneath them.
Can the model tell you apart from organisations that share your name, acronym, or category? This is where we see the most surprising failures. A UK business getting confused with an offshore drilling company. An energy company getting blended with an unrelated overseas business. A consultancy whose acronym matches a regulator’s. Disambiguation is the single biggest source of “the AI got it completely wrong” in the work we’ve done.
Does the answer reflect the current state of affairs, or is it stuck a few years back? Old leadership, old pricing, missing recent openings, expired partnerships. Grounded mode usually fixes this. Training mode almost never does.
If a user feeds in a false premise (“I heard you were acquired by X” / “didn’t you close all your London sites in 2023?”), does the model push back, hedge, or quietly agree? This one’s tricky. Stakeholders, journalists and prospects may already be arriving at AI tools having heard something incorrect. A good model corrects them but a bad one could reinforce it.
Can it discuss your brand beyond surface level facts? Your sustainability commitments, your specific competitive advantage, your customer experience, your real positioning? Or does it default to the press release version of you?
Aleyda Solis calls the combined score “representation accuracy,” and it’s the one AI visibility measure that tells you whether you’re being understood, rather than just mentioned.
One of our clients, a well-known restaurant group, was told by one model, answering from memory, that it didn’t hold a certification it’s had for years. With thin data to go on, the model guessed and said it with confidence. Grounded mode, where the model searches before answering, got it right.
Seven months later, on newer models, that error had gone. Same with ownership at an energy company we work with. In April every model got the owner wrong from memory. By September ChatGPT had it right, even though the change happened years ago. Depending on the industry, we’ve seen ChatGPT’s training data catch up five to seven months after a fact is well covered.
The other models were still guessing, and not consistently. One gave two different wrong owners depending on how we worded the question. Another turned a well known founder’s surname into one that doesn’t exist. Some models say they don’t know. Others pick something plausible and commit.
The same energy company got nearly everything right when models searched, with one exception. Ask about its overseas work and every model returns a different company that dominates search results in that region. This is the rare case that content can’t fix on its own. It needs disambiguation pages, schema, an llms.txt file that says who you aren’t, and keeping hold of old domains. The disambiguation page we built was cited 15 times in September, and ChatGPT now gets the overseas question right. Big brands get caught too: one model mixed up a well known consumer brand with a finance firm of the same name.
Sometimes the problem is on the site. One brand quoted two versions of its headline number, one on the About page and one in a press release. One model doubled it. Another found both and hedged.
Being known and being recommended are different things. One client answered fine when asked about itself, but never appeared when we asked which businesses in its sector were best. Five months later it was in most of those answers, top of the list in one. Most unprompted mentions come from third-party pages. AirOps puts it at 85%.
Models handle false claims very differently. ChatGPT, answering from memory, went along with two of eight false premises in one audit, including one that got the brand’s business model backwards. Perplexity, which always searches, rejected all eight. A brand with years of press coverage saw every false claim rejected by every model. A younger one with thinner coverage didn’t.
Grounded answers catch up on news fast. A restaurant group opened a new site in September and three days later every model that searched had it, date included. One answering from memory put the same opening two years earlier.
Training data goes stale in quieter ways too. One membership organisation’s founding story was perfect, but the models couldn’t name its current leadership or anything it had published in the last two years. That gets fixed off site: Wikipedia, news, policy databases, the places the next training run will read.
When a brand has a framework (pillars, promises, values) but no clear page listing it, models make one up. Earlier this year only Google’s AI Overview got one brand’s values right, from a page on a secondary domain. By September three of four grounded models had them. From memory, one was still reciting a set that doesn’t exist.
Findings split roughly two ways, and they need different fixes.
Training data gaps are long cycle work. You’re influencing what the next generation of models ingests, which means content on your own site, but also third party authority: Wikipedia, news coverage, industry databases, structured data that ends up in the open knowledge graph. You won’t see the impact for months, but this will yield long-lasting, consistent results.
Grounded gaps are short cycles. They’re almost always fixable on the website itself: missing canonical pages, weak schema, no structured source for the thing the model is being asked about, current information that exists in a PDF. We can see these changes make a difference within weeks.
The audit isn’t the deliverable. What comes out of the audit is a prioritised list: which content is missing or out of date, what needs looking at technically, and which third parties we should be talking to.