Ask ChatGPT for the best dog food subscription, a decent cordless vacuum or a skincare brand that won’t set your face on fire, and you get a shortlist. Four or five names, a line on each, done. For a lot of people that shortlist is now where the decision starts, and increasingly where it ends. Kevin Indig’s usability study from April 2026 watched people shop in Google’s AI Mode and found about three quarters of their final shortlists came straight from the AI’s output, with no second opinion and no checking elsewhere. Miss the list and you’ve probably missed the sale.
Those lists aren’t random, though, and you can’t buy your way on. They’re built from what the web already says about you. Which means you can change them, though it takes time and real work rather than a plugin.
Why AI recommends some brands and not others
When an AI tool answers a buying question it leans on two things: what it learned in training, and what it finds when it goes and searches the web live (which most of them now do). In both cases it’s looking for evidence that a brand is credible and relevant to the question. It wants proof.
The strongest signals, from running our own checks and from the research that’s actually been published rather than the stuff people say on LinkedIn:
Coverage in publications the model trusts. Muck Rack’s May 2026 study looked at millions of links cited in AI answers and found more than four in five come from earned media, with journalism alone about a quarter. Press coverage is most of the raw material.
Mentions count as well as links. Ahrefs looked at tens of thousands of brands and found web mentions track AI Overview visibility about three times more closely than backlinks do. Being talked about beats being linked to, by some distance.
Consistent descriptions. When 10 independent sources describe you the same way (“the subscription for fussy dogs”), the model repeats it with confidence. When they all say something different, it hedges, or skips you.
Pages on your own site that are easy to quote. Plain answers to real questions. We’ll come to that.
Real conversation. Reddit, review sites, forums. Pew Research found Wikipedia, YouTube and Reddit were the most cited sources in Google’s AI summaries. People talking about you honestly, in public, counts.
Notice what’s missing: your ad budget, your follower count and your page-three rankings. The answer isn’t built from any of them.
Step 1: find out where you stand
Don’t start with “what is [brand]?”. The model knows who you are, that’s not the problem. Start with the questions people type before they know you exist. “Best eco cleaning products UK.” “Which savings app actually pays a decent rate.” “Is [competitor] any good.” Write down 20 of them. Real ones, in the words a customer would use, not the words your brand guidelines would use.
Run each one in ChatGPT, Gemini, Perplexity and Google (look at the AI Overview and AI Mode). For every answer note three things: are you named, how are you described, and who’s named instead of you. That’s your baseline share of answers. It’s usually a bit grim the first time, and it usually shows a pattern within about 10 minutes. One competitor everywhere. One publication quoted over and over. A description of you that’s three years out of date.
Step 2: learn which sources AI trusts in your category
Perplexity, Google’s AI surfaces and ChatGPT with search all show their sources. Collect them. Every one, in a spreadsheet, for all 20 questions. Within a few dozen answers you’ll see the same places turning up again and again: a couple of national newspapers, a specialist magazine or two, a review site, a Reddit community, maybe one blogger who’s somehow become the oracle for your category.
That list is your real media list. And I’ll be honest, it’s often nothing like the one a traditional PR plan would start with, because it’s ranked by influence on AI answers rather than by circulation or domain authority. Muck Rack’s December 2025 report found that when you compare the journalists brands pitch most with the journalists AI engines actually cite for that brand, the overlap is tiny, a couple of per cent. Most PR effort is pointed at the wrong people for this particular job.
The other thing you’ll notice is that the sources aren’t necessarily ranking well. Ahrefs looked at millions of AI Overview citations in March 2026 and fewer than four in 10 were in the top 10 results for the query. So nearly two-thirds of what AI Overviews cite sits outside the first page, and “we rank, therefore we’re fine” isn’t a safe assumption any more.
Step 3: earn coverage in those sources
This is where digital PR does the heavy lifting, and it’s the bit no tool can shortcut. The aim is to be mentioned, described accurately and ideally recommended in the places you found in step 2. The tactics aren’t exotic. Data-led campaigns that give a journalist something worth writing about, built on your own data or new research. Reactive PR, getting your experts quoted on the news breaking today, because fresh coverage gets cited (Muck Rack found more than half of dated journalism citations were from the last year). Product placements in the round-ups and “best of” lists those publications publish. And expert positioning, so your founder or your specialist becomes the person a journalist rings.
One thing I’d push hard on: be precise about how you’re described. The words a journalist uses are the words the model will repeat, possibly for years. “A dog food brand” and “the subscription that tailors each dog’s food to its breed, age and weight” are not the same sentence to an AI. Make every placement work twice.
Step 4: make your own site easy to quote
AI tools still read your site. A lot of brands make that really hard. Say plainly what you are, who you’re for and what makes you different, near the top of your key pages, not buried under a hero video and three paragraphs of “we believe”. Answer the questions customers actually ask in short, self-contained paragraphs that would make sense lifted out on their own, because that’s exactly what happens to them. Use proper headings. Add structured data (product, FAQ and organisation schema) so the facts are machine-readable. Keep facts consistent everywhere: prices, founding year, locations, awards.
And check your robots.txt. I’ve lost count of the sites that block the AI crawlers by accident and then wonder why ChatGPT thinks they went bust in 2022. There’s no trick here: write so people understand you, then sort out the plumbing.
Step 5: measure share of answers as well as rankings
Re-run your 20 questions every month. Same questions, same tools. Track share of answers, how you’re described and which of your placements are being cited. Then add AI tools as a source in your analytics, because ChatGPT, Perplexity and Gemini all send referral traffic you can see, and it’s good traffic: Adobe’s UK data for August 2026 showed visitors from AI sources converting about a fifth better than other channels.
But the bigger effect is the one your dashboard won’t attribute. Similarweb found that being recommended by an AI makes someone more than twice as likely to visit your site in the following week, and most of that arrives as branded search, not as an AI referral. So watch branded search too. If coverage is landing and the answers are changing, branded search moves before anything else does.
Over a few months that tells you something rankings never could: whether the coverage you’re earning is changing what customers hear when they ask for advice. And right now the window’s open. Semrush looked at more than a thousand ChatGPT categories in the first half of 2026 and found only about one in seven had a clear owner. The rest are up for grabs. I’d get on with it.