Ask them. That’s the whole trick, and it costs nothing: open ChatGPT, then Perplexity, then Gemini, and type in the question one of your customers would type. Not “accounting services.” Something like “who’s good for bookkeeping if I freelance and live in Austin.” Fifteen or twenty phrasings of that shape. Jot down whether your name turns up. Congratulations, you now have a baseline for your AI search visibility, and it took less time than a stand-up meeting.
Second half, and people skip it constantly: go look at your own logs. Filter for GPTBot, ClaudeBot, PerplexityBot. Check whether chatgpt.com or perplexity.ai ever appear as a referrer in your analytics. Prompts tell you what the machines say about your business. Logs tell you whether they’ve ever read a word of it. One without the other will mislead you.
Right. The longer version.
Why anyone bothers measuring this now
Maybe you’ve had the odd experience where the rankings look fine and the traffic doesn’t. Not a cliff. Just a slow leak nobody can locate.
Cloudflare put a number on part of that. They publish a thing called the crawl-to-refer ratio: pages an AI platform grabs off your site, divided by visitors it sends you back. For the week of 19 to 26 June 2025 the published table ran from Anthropic at roughly 70,900 crawls per single referral all the way down to Mistral, which sat at 0.1 to 1 and was therefore sending ten times more people than it took pages. Cloudflare’s own read on the trend, in their words, was more crawls and fewer referrals (Cloudflare Blog, 1 July 2025).
Strip the engineering out and it means this. Your stuff gets eaten and reassembled into an answer. Whether your name survives the reassembly is now a completely separate question from whether you rank, and it needs its own measurement rather than a bullet at the bottom of somebody’s monthly report.
Step one: write the prompts like a customer, not like you
Almost everyone gets this wrong on the first try. They type their own brand name in, see it come back, feel relieved, close the tab. Well, yes. You asked about yourself.
What counts is the question asked by somebody who hasn’t picked yet. Aim for fifteen to twenty-five prompts spread over roughly four shapes:
- Category. “Best commercial cleaning companies in Manchester?”
- Problem. “My warehouse stock counts keep drifting out of sync, what software actually fixes that”
- Head to head. “Competitor A or Competitor B for a team of twelve”
- Qualified. “Dental implants in Bogotá, financing available”
Use their vocabulary. A restaurant owner hunting a supplier types “where to buy bulk olive oil for restaurants.” Nobody has ever typed “wholesale procurement solutions” into anything, ever. If the list you’ve written sounds like a slide from your own pitch deck, bin it and start over.
Step two: run the identical list through each engine
Test one and extrapolate to the rest and you’ll get burned. They pull from different sources and weight different signals, and the spread is wider than seems reasonable. I’ve watched a company come up in four out of five Perplexity answers and score a flat zero in ChatGPT on the exact same wording. Same day. Same question.
Quick notes per engine
ChatGPT. Fresh chat, browsing switched on, and kill memory or personalisation if the setting is available to you. Skip that and you’re measuring your own browsing history wearing an engine’s clothes.
Perplexity. The friendly one, because it stacks numbered citations right underneath. Pay attention to which domains it leaned on, not merely whether it said your name.
Google AI Overviews. Logged out, incognito, and set to the place you actually sell in. Location swings these results far harder than most people expect.
Gemini, Claude. Do a pass if the afternoon allows. Smaller volumes today. Same machinery underneath.
Two or three runs per prompt, minimum. These things are non-deterministic and one result is an anecdote, not a measurement.
Step three: an extremely boring spreadsheet
Row per prompt. Column per engine. In each cell, three facts: mentioned or not, linked or not, and which rivals showed up.
Out of that fall two numbers. Mention rate, the share of prompts where your brand appears at all. Citation rate, the tighter one, where an engine links to your site as an actual source. Citation rate is harder to move and worth considerably more.
Zero on both? Completely ordinary for a first pass, and not worth losing an evening over. The number that matters is the one you get in thirty days, then sixty. Rerun the same list monthly, same prompts, same conditions, or you’re comparing nothing to nothing.
Step four: make sure they can read you in the first place
Here’s the check that gets skipped, and it explains a startling share of the zeros.
Open your server logs, or Cloudflare’s analytics if that’s where you live, and filter on AI user agents. The ones worth hunting: GPTBot and OAI-SearchBot from OpenAI, ClaudeBot from Anthropic, PerplexityBot, and Google-Extended. Cloudflare keeps a public directory of verified bots if you need to check what some unfamiliar string belongs to.
Then read your robots.txt. Properly, line by line. I have genuinely lost count of the sites quietly blocking GPTBot because somebody pasted in a block-the-scrapers snippet back in 2023 and no human has looked at the file since. Deciding to block training crawlers is a legitimate business choice. Doing it by accident while wondering why ChatGPT has never heard of you is a different situation entirely.
One more, less obvious: does your content exist in the HTML, or does JavaScript assemble it after the page loads? A crawler that doesn’t run scripts gets an empty shell. Product descriptions living inside a component that only populates client-side make you invisible, and nobody involved ever intended that.
Step five: look hard at whoever got cited instead
This is the part that tells you what to go build, so don’t rush past it.
Read the sources the engines did cite. Frequently they aren’t competitor homepages at all. A Reddit thread. Some comparison piece on a trade publication. A G2 listing, a local directory, the transcript of a podcast episode nobody remembers recording. Engines lean on third-party context for a fairly sensible reason: a company describing itself is the least reliable witness available on the subject of that company.
Which means the fix usually isn’t another page on your website. It’s turning up, accurately, in the places these systems already trust. Fix the directory entry that keeps surfacing. Answer questions properly in the forum where your category gets argued about. Get yourself into the roundup article that owns your category term. Deeply unglamorous work, all of it. Moves AI search visibility more than another homepage rewrite ever will.
So the score is zero. Now what
Assuming crawlers can actually reach you, four things tend to work, roughly ordered by how much effort they cost:
- Answer in the opening paragraph. Not after four hundred words on the history of your sector. These systems lift passages that stand up alone, so hand them one.
- Publish what only you know. Actual prices. Actual timelines. Numbers pulled from your own jobs. Generic advice is already covered by a thousand pages the model has swallowed; a specific figure with something behind it is quotable.
- Be one consistent entity. Identical name, address and description everywhere, plus structured data on the pages that matter. Models are doing entity resolution and every inconsistency makes you fuzzier.
- Earn mentions off-site. Back to step five. Slowest lever. Usually the one that works.
When you should just hire someone
Honestly? Often. Measuring is a proper DIY job, and the routine above fits neatly into one afternoon with a coffee. Acting on what it tells you is a different animal.
Competitive category, multiple countries or languages, an existing SEO programme this has to slot into, or a fix that means six months of content and digital PR? A good agency or a seasoned freelancer will get you there faster than you’ll get yourself there, and it isn’t close. They’ve watched this pattern play out across dozens of sites and they know which lever tends to shift first. Paying for that instinct is nearly always cheaper than buying it with your own hours.
The thing not to do is hire before you’ve measured. Run the baseline first. Then you’re briefing from evidence, and there’s a number on record that existed before anybody started billing you.
Where Hepteon comes into this
All of the above is a loop. Measure, read, fix, measure again. It works fine. It’s also precisely the sort of loop that quietly dies around week three, because you have a business to run and the spreadsheet does not chase you.
Hepteon exists for that reason. Seven AI agents run a website end to end, named for the seven events of the heptathlon: Strategist, Connector, Technical, Writer, Amplifier, Results and Publisher. Between them they watch how the site shows up across search and AI engines, catch whatever is blocking the crawlers, write what fills the gap, and report what genuinely moved against whichever goal the owner set.
Agencies and freelancers can run the whole thing white label for their own clients, which is the same loop at portfolio scale instead of one site at a time.
Start with the spreadsheet though. Twenty prompts, three engines, one afternoon. Whatever you decide afterwards, you’ll be deciding it on numbers rather than a feeling.
Featured image: “Woman using laptop while holding a cup of coffee” by homegets.com, licensed under CC BY 2.0.
Once you know whether AI tools mention you, the next question is where to spend your effort. That is exactly what SEO vs GEO vs AEO: What Actually Changes for Your Business breaks down.
If you have already noticed fewer visits and suspect AI is pulling them away, see Why Your Website Traffic Dropped — and How to Tell If AI Is the Reason.