AI & Technology

AI has an opinion of your brand. Are you influencing it?

By Benjamin Knapp, Executive Director, Lab, Saffron Brand Consultants

Nine in ten marketing leaders call AI visibility a priority. Almost nobody measures it. 

Ask ChatGPT the questions your buyers ask. Not “tell me about us”, but the real ones: who should I shortlist, is this company any good, what are the alternatives. You’ll get a fluent, confident answer. It might be right. It often isn’t. Either way, it’s being repeated thousands of times a day to people who never show up in your analytics. 

This isn’t fringe behaviour any more. 

Ofcom says 54% of UK adults now use tools like ChatGPT or Gemini, up from 31% in 2024, and about 30% of UK Google searches now return an AI summary. In the US, SparkToro puts the share of Google searches ending without a click at 68%. Not all of that is AI, but the direction is obvious. The answer is replacing the visit. 

Marketers know. 

Forrester finds nine in ten B2B marketing leaders now treat AI visibility as an investment priority. So, the gap isn’t awareness. It’s that hardly anyone has turned the interest into a routine. Most teams are at the screenshot stage: a pilot here, an odd ChatGPT answer forwarded on Slack there. We measured human perception this way once too, before trackers existed. 

The missing step is to treat AI systems as an audience in their own right. Customers, employees, investors and regulators each have an owner and a tracker. The assistants now answering on your behalf have neither. They don’t fill in surveys or sit in focus groups; they form their view from whatever the public record gives them, and they repeat it to anyone who asks. 

Don’t rush to buy a dashboard 

The obvious move is to buy one of the new AI visibility dashboards. I’d wait. Scores vary between tools and can shift from one run to the next, because the systems themselves rarely give the same answer twice. A tidy number means very little if you can’t see what sits behind it. 

Some of these tools are built by serious people and will get better. Right now, outsourcing your understanding to them isn’t measurement. It’s comfort. 

A brand tracker for the answers nobody owns 

What works is duller. Treat it like brand tracking: a few stable measures, tracked quarterly, benchmarked against competitors, owned by whoever owns your tracker today. 

Five measures do most of the work: 

  • Accuracy: does the answer get the basic facts right? 
  • Recommendation: when someone asks for options in your category without naming you, do you come up, and where? That’s unaided awareness, machine version. 
  • Framing: what story does the answer tell? Leader, challenger, has-been? 
  • Consistency: do ChatGPT, Gemini, Claude and Perplexity say roughly the same thing? If they don’t, your public record is thin enough to be read several ways. 
  • Sources: which pages, reviews and databases are the answers actually built on? 

The method is deliberately boring. 

Write down the 20 questions your buyers actually ask. 

Put them to the four main assistants, three times each, because the answers vary. Score the results against the record, blind, two people. Do the same for three competitors. Repeat every quarter. It’s about two analyst-days per wave, and the corridor anecdote becomes a number with a trend line and an owner. 

One airline, four different answers 

We ran a cut-down version on easyJet on 13 August, the week it agreed a £5.7bn takeover by Apollo. Two of the four assistants missed the deal entirely. One of those also repeated a passenger milestone the airline’s own numbers contradict, traceable to one aggregator page. 

The other mixed divisional revenue up with group revenue and produced a profit figure that appears nowhere in easyJet’s reported results, directly under quarterly numbers that were accurate to the company’s own release. A third caught the deal but skipped the record year before it. One of four got everything right. Same day, same questions, same public record, and nothing in any single answer tells you which you’ve got. 

You can’t edit the model. Start with the record. 

The sources measure is the lever. You can’t edit a model, but you can edit the record it reads. If the answers lean on stale press releases and a Wikipedia page nobody has touched since 2021, that’s fixable, and the fix looks familiar: corrections, fresher third-party coverage, the same facts stated the same way everywhere. People have started calling this machine-readable reputation. 

Call it what you like. The practice matters more than the label. 

AI has already formed its opinion of your brand, from whatever evidence it found lying around. Changing that opinion isn’t a dark art. It’s housekeeping, measured properly. Do that, and you’ll probably never need the panic currently being sold to you. 

Related Articles

Back to top button