Nobody clicks the link

Every AI-visibility pitch ends with the same promise: get cited, get traffic. The click-through from an AI answer has now been measured. It is close to zero — and that is not the reason to walk away.

It is the most common objection we get, and it usually arrives a few weeks after the work is done. The business is being named now. Ask ChatGPT for a tax adviser in their town and there they are, second in the list, with a link next to the name. And the analytics dashboard shows nothing. No spike, no referrals worth the name, the same quiet week as before. So what exactly did we buy?

It is a fair question. The honest answer is that a citation and a click are two different events, and only one of them was ever likely.

The click-through, measured

Pew tracked the real browsing of 900 US adults for a month and compared what happened after searches with and without an AI summary (Pew Research Center). Seeing a summary roughly halved the chance of a click going anywhere. It also raised the share of sessions that simply ended on the spot, from 16% to 26%. The summary answered the question. There was nothing left to click.

The crawler side of the ledger reads worse. Cloudflare classified a year of AI crawling by purpose: 80% of it was for training, 18% for search, and 2% for user actions — and only that last sliver can end in a link a human follows. Set the crawling against the referrals and you get the gap in one number. In July 2025, for every visitor Anthropic's systems sent back to a website, its crawlers had already requested about 38,000 pages. OpenAI's ratio was 1,091 to one, Perplexity's 195 to one (Cloudflare).

Two caveats, because they cut against the number. Anthropic's ratio had fallen 86.7% in the six months to July 2025, so the gap is closing, not widening. And Cloudflare counts referrals by the referrer header, which native apps do not send — so the true ratios are smaller than the published ones, by an amount nobody can put a figure on (Cloudflare Radar). The order of magnitude survives both. AI search is not a traffic channel in the way Google was a traffic channel.

So what does a citation actually buy?

A place on the shortlist. That is all — and it turns out to be worth a lot, because of what people do next.

BrightLocal put the question to 1,002 US consumers: 45% had asked an AI tool for a local business recommendation in the past twelve months, up from 6% a year earlier. That makes AI the third most used source for local recommendations, behind Google and Facebook and ahead of Yelp and Tripadvisor. Among the 455 who had used it, 88% check what they are told — 51% to see whether a review is genuine, 37% to see the source — and 97% at least sometimes hold the recommendation up against real reviews (BrightLocal).

That mechanism also explains the empty dashboard. The person did not click your link. They read your name in an answer, opened a new tab, searched for the firm, read four reviews and called. In your numbers that is a direct visit or a branded search — the same two lines as always. The citation did its work in a place you were not looking.

The few clicks that do arrive are unusually good ones. Semrush measured the average AI-search visitor at 4.4 times the value of an average organic search visitor, on conversion rate, on the grounds that someone arriving from an AI answer has already compared the options (Semrush). Read it with its limits: it covers 500-odd digital marketing and SEO topics, which is Semrush's own market, and a conversion-rate ratio is not a revenue promise.

What to measure instead

If you have five employees, the practical consequence is that a traffic graph will mislead you twice over. It will show nothing while you are winning, and it will not show the loss when you quietly drop out of the answers.

Measure the mention. Take the ten or fifteen questions somebody actually asks before hiring a firm like yours, put them to ChatGPT, Google AI Mode, Perplexity and Claude, write down how often your name comes back, and run the same set again every few weeks — a single run is noise, for reasons we have been through. That count moves before anything else does.

Then put the remaining effort into the second half, the part nearly everyone skips. Nine in ten of the people who hear your name go and check it. A review profile with eleven entries, the newest from two years ago, is where a citation goes to die. Fixing that is a fortnight of asking customers, not a quarter of outreach.

And the honest limit: if nobody in your category asks an AI anything, none of this earns you an afternoon. That is the first thing to test, not the last.