AI Visibility: How to Measure It and Improve It

A prospect emailed me in July with one line: "We don't exist in ChatGPT. Can you fix it?"

Fair complaint. Terrible brief. When I asked what they had actually checked, it turned out the founder had typed the company name into ChatGPT once, got a mushy answer, and panicked. No prompt list. No server logs. No idea whether an assistant could even read the site.

That's roughly where most companies sit. The feeling of being invisible is real; the measurement is missing. And you can't improve a number you never wrote down. So let's write it down.

What AI visibility actually means

Two separate things get squashed into one phrase, and mixing them up is why so many "AI visibility" projects go nowhere.

The first half is being mentioned. Someone asks Claude, ChatGPT, or Perplexity a question your product answers. Does your brand come up? Does your domain land in the citation list? This half looks a lot like SEO because it mostly is SEO. Google's own AI optimization guide says it plainly: "optimizing for generative AI search is optimizing for the search experience, and thus still SEO."

The second half is being readable. Can a crawler, a fetcher, or a browser agent open your page, work out what's on it, and finish a task there? That has nothing to do with content strategy and everything to do with how your HTML is built.

The two are linked in an irritating way. If nothing can read your page, nothing can cite it. But readability earns you no mentions on its own. Both halves, measured separately.

Why anyone should care: the Shopify numbers

The evidence I keep coming back to is platform-wide rather than one flattering case study. Shopify reported for Q1 2026 that AI-driven traffic to its merchants grew more than 8x year over year, and orders from AI-powered search roughly 13x. The quality figures are the interesting part:

• AI-referred shoppers convert about 50% higher
• Average order value runs 14% higher
• AI referrals beat organic in 23 of 25 categories
• Over half of AI-referred sessions land straight on a product page, against about 20% for organic

That last line is the one to sit with. Assistants skip your homepage. They drop the buyer on the exact page that answers the question, which quietly turns every product and service page into a landing page whether it was designed as one or not.

Now the counterweight, because I'd rather you plan than panic. TollBit's H1 2026 publisher data put AI apps at 0.05% of external referrals for EU publishers and 0.16% in North America, with Google at around 84%. Small, high-converting, growing fast. Get ahead of it, but don't expect it to pay this quarter's rent.

How to measure AI visibility, hands on

Four checks. All free. All doable this afternoon.

1. Ask the assistants your buyers' questions

The boring one everybody skips, and the only one that measures mentions.

Write down 10 to 15 questions your buyers actually ask. Not "best CRM software" but the way a human types mid-problem: "how do I move patient records off an old EHR without downtime." Run each through ChatGPT, Claude, Perplexity, and Gemini, and record three columns: were you mentioned, which domains got cited, which competitor took your slot.

Answers are non-deterministic, so run each question a few times and treat the result as a sample. The cited-domains column is the most valuable thing in the sheet: it tells you which third-party pages assistants trust in your category, which makes it your outreach plan.

Repeat monthly with the same prompts. An hour a month, and it's the only AI visibility metric that maps to revenue. For the deeper version, we broke down what earns citations in LLM SEO: what actually gets you cited.

2. See what an agent sees

Most AI crawlers don't render your page the way your browser does, so go look at your site the way they do.

Paste `https://r.jina.ai/` in front of any URL. Jina Reader hands back the page as plain markdown, free, up to 20 requests a minute. Do your three most important pages and check one thing: is the content there? A shell with navigation and no body copy means you found your problem before spending a cent.

3. Run Lighthouse's Agentic Browsing audit

Google added an `agentic-browsing` category to Lighthouse in v13.3.0 (May 2026). It ships in Chrome 150+ DevTools and PageSpeed Insights, and runs from a terminal:

npx lighthouse https://yoursite.com --only-categories=agentic-browsing

The score is a fraction rather than a 0–100 grade: Google's position is that "the current focus is to gather data and provide actionable signals rather than a definitive ranking." Six audits exist on paper; three cover WebMCP and return N/A on standard Chrome, so you're really taking a three-check test: the accessibility tree (38 axe rules, one violation drops you to zero), cumulative layout shift, and llms.txt format.

Read that last one carefully. The audit checks the file's format, not whether you have one. No file at all scores fine; a malformed file scores zero. A bad llms.txt is worse than none.

What we measured on 2026-08-26 with Lighthouse 13.4.1:

Yes, we failed our own audit on the first run. Rewriting our llms.txt into proper markdown links got us back to 3/3 the same week. Shopify and Airbnb failed on the accessibility tree. If engineering teams at that scale drop this check, yours probably does too. We walked through the internals in what Lighthouse's Agentic Browsing audit actually checks.

4. Run Cloudflare's Is It Agent Ready

Cloudflare put a free public checker at isitagentready.com. No signup. It grades discoverability, markdown negotiation, bot access control, protocol discovery (MCP Server Card, WebMCP, API catalog), and commerce protocols. The same grading appears in Cloudflare Radar's URL Scanner.

Expect a low score, then relax. Cloudflare's scan of the top 200,000 domains found 78% have a robots.txt, 4% declare Content Signals, 3.9% pass markdown negotiation, and fewer than 15 sites in the entire dataset publish both an MCP Server Card and an API catalog. Read that as headroom.

Then read your own logs, segmented by user agent (GPTBot, ClaudeBot, OAI-SearchBot, Claude-SearchBot, PerplexityBot, ChatGPT-User) and by status code rather than raw counts. One attribution warning: the Claude app sends no referrer, so any AI-referral number you produce is a floor. Our complete guide to AI crawlers covers what each agent is for.

Why sites end up invisible in AI answers

Three causes, in the order I usually find them.

Your content only exists after JavaScript runs

Vercel's crawler analysis (November 2024 data) found that the major AI crawlers from OpenAI, Anthropic, Meta, and Perplexity fetch JavaScript files but never execute them. A client-side-rendered page hands them an empty shell. Nothing to quote, so nothing gets quoted. Worth flagging that this argument comes from Vercel; Google is more relaxed about code quality.

The same analysis found AI crawlers burn roughly a third of their requests on 404s: 34.8% for ChatGPT's crawler and 34.2% for Claude's, against 8.2% for Googlebot. They're guessing at URLs because your sitemap and internal links don't say where anything lives.

Your markup doesn't say what things are

Google's web.dev guidance describes three ways agents perceive a page: screenshots, raw HTML, and the accessibility tree, which it calls "a high-fidelity map that ignores the visual 'noise' of CSS." A `<div>` with a click handler doesn't appear on that map. Neither does an unlabeled input, or an icon-only button. Adobe reported that 80% of its LLM Optimizer early-access customers had critical content-visibility gaps, which matches what I find in audits.

You're accidentally blocking the bots that matter

Every vendor runs a search bot you have to allow if you want to appear at all. OpenAI's documentation is explicit that sites opted out of OAI-SearchBot "will not be shown in ChatGPT search answers." Blocking that is self-inflicted invisibility, and a different decision from blocking GPTBot, which is about training data.

And robots.txt is a preference, nothing more. TollBit measured around 15% of AI scrapes ignoring an active disallow in H1 2026, with ChatGPT-User at 54%. Real enforcement lives at your CDN or WAF.

How to improve AI visibility, in priority order

Ranked by strength of evidence rather than novelty.

On llms.txt, the picture is messy. Google says outright that Search doesn't use these files and that publishing one "will neither harm nor help your site's visibility or rankings in Google Search." Ahrefs studied 137,210 domains and found 97% of llms.txt files got zero traffic in May 2026. Yet every major AI lab publishes one and Lighthouse checks its format. A developer-docs convention that got mistaken for a search lever; we went through the data in does llms.txt actually work.

And notice that priorities 1 through 5 are all just competent web development. Google's closing line on agent-friendly sites is that "everything we suggest to make a site 'agent-ready' also makes sites better for humans." The accessibility work your team deferred for six years now has a second business case, which is a strange kind of vindication. Same argument we've made about user experience in mobile app development: unglamorous fundamentals win.

Where to start on Monday

Run the Jina Reader check on three pages, then Lighthouse on one. Fifteen minutes. If the markdown comes back empty, close every tab about llms.txt and go fix your rendering, because nothing further down the list matters until a machine can read the page. Then open the spreadsheet: ten questions, four assistants, once a month. In three months it'll be the only data showing whether any of this worked.

If you'd rather someone else find the failures, we run a free GEO audit: we score your pages, list what's blocking assistants, and rank the fixes. What you do with the list is your call.

Vladimir Terekhov

Co-founder and CEO at Attract Group