The 14% problem: why a single AI visibility score is a category error

Profound analysed 680 million citations across ChatGPT, Perplexity and Google AI Overviews to map what each engine actually cites. BrightEdge then measured how much five AI engines' citation lists overlap: at best 59%, at worst 16%. And Google's AI Mode (the full chat surface, not the AI Overviews that sit on top of normal results) shares just 14% of its URLs with Google's own top ten organic results, per SE Ranking.

It’s a pretty big problem when Google's AI doesn't even agree with Google.

If you're old enough to remember search before Google ate it, this is familiar territory. AltaVista, Lycos, Yahoo, Ask Jeeves and Hotbot. Different surfaces, different playbooks, different things that worked. We forgot about all of that because Google won and the optimisation job rapidly collapsed into one engine, one SERP, one set of rules. For about fifteen years, "SEO" meant "Google." They had a pretty good run.

Now we're back in 1998 again, except the engines are heavier, the playbooks are less obvious and the people selling you dashboards are acting like nothing’s changed.

What ChatGPT, Perplexity and Google AI Overviews each cite

Same query, three different answers. Profound's 680-million-citation dataset (August 2024 to June 2025) shows the engines don't share a citation graph. They barely share a vocabulary. BrightEdge's April 2026 follow-up puts numbers on the gap: pairwise citation overlap between engines runs from 59% down to 16%. (BrightEdge's own headline is that the engines often land on similar brands via different sources. That's not a comfort. The sources are the part you can influence.)

The sharpest single shift Profound has tracked: LinkedIn climbed from around #11 to #5 in ChatGPT's most-cited domains in three months, November 2025 to February 2026, with its citation frequency more than doubling. The entity graph moved, fast.

The click side has numbers too. When an AI Overview is present, organic click-through drops 61%, from 1.76% to 0.61%: that's Seer Interactive's data across 3,119 informational queries and 42 organisations, June 2024 to September 2025. Being the cited brand claws some of it back: brands cited in the AI Overview see 35% higher organic click-through than brands that aren't. Getting cited isn't a vanity metric because it's where the actual clicks went.

(If you don't know what your structured data looks like to a machine, the Schema Sniffer will show you in about ten seconds.)

One brand can be cited by ChatGPT and invisible to Perplexity. Another can dominate Perplexity and never appear in a Google AI Overview. This isn’t mere nuance between models, it's the shape of the new search layer.

The Reddit paradox

Reddit is trusted by the models precisely because it's hard to manipulate and this is where the industry's been telling on itself.

Reddit gets cited by all three engines. Perplexity leans on it hard. Google AI Overviews lean on it hard. ChatGPT less so, but it's still there. So naturally there's been a flood of "how to manipulate Reddit for AI visibility" articles from self-appointed experts in the last eighteen months. What's remarkable is how many of those articles are admitting the opposite of what they claim.

The downvotes, the mods, the self-policing community norms: they're the same kind of friction that made .edu and .ac.uk domains trusted in the days we were all trying to manipulate the original PageRank. The assumption was that spammers couldn't easily game university domains, so a link from one carried weight. Reddit works the same way. The platform is designed to surface consensus and bury bad-faith contributions, which is exactly why the AI models love it so much. The manipulation guides exist because it's hard to manipulate. The difficulty is the whole point and the people writing the guides are mostly just describing the wall they hit.

Yes, of course Reddit gets gamed. So did .edu, or even .gov domains. The point is the gaming is the exception, not the rule, and the models are weighting it on the assumption that it mostly isn't. The "just go manipulate Reddit" playbook is the AI-visibility equivalent of "just buy .edu links" in 2006. Some people will. Most will get caught, downvoted and end up worse off than if they'd done nothing.

The category error

So how can this be measured? A single AI visibility score is a category error: it averages engines that can disagree on five citations out of six.

The same vendors publishing the citation studies will sell you a dashboard that compresses ChatGPT, Claude, Perplexity, Gemini and Google AI Overviews into one number. Goes up, goes down. Dashboard-friendly. Easy to put in a stakeholder deck.

It's the same instinct that gave us PageRank-as-one-number, Domain Authority-as-one-number, "your ranking" as one number. The industry has spent twenty years learning to sell complexity by condensing it into a single figure that moves up or down, and in the new frontier the AI visibility vendors are running the same play, because it's the only packaging they know how to sell.

But the data they're publishing themselves is the refutation. If two engines can agree on as little as one citation in six, a single score is a map of London that only shows Oxford Street. It's not wrong, exactly. It's just not telling you where you are. A brand can have a great "AI visibility score" and be invisible to Perplexity, or strong on ChatGPT and absent from the Google AI Overviews where its customers are asking questions.

I know how these numbers travel, because this post nearly shipped with a bad one. An earlier draft led with a tidy stat: only 12% of domains are cited by both ChatGPT and Perplexity, attributed to Profound. Except on closer inspection, this number isn't in Profound's study. As far as I can trace it, it exists only in aggregator blogs citing each other, each attributing it to a different vendor, with no original source. A made-up number about citation integrity, laundered through the exact ecosystem the models are trying to filter. The single-score problem and the sloppy-stat problem are the same problem: numbers that travel better than their sources.

The AI visibility platform pedlars will have to fix this. They probably know it already: the methodology pages admit as much. 5W's own AI Visibility Index discloses that its figures are "directional estimates", "not the output of logged query runs" and "intended as a strategic framework — not a definitive search engine measurement". This is at least honest but it’s also slightly at odds with selling it as a measurement product. The next generation of these tools will have to be engine-aware, or they'll be sold to people who don't know the difference, which is a smaller market than it looks.

5W also claims more than a third of consumers now begin product research with AI, not Google. Treat that one as directional too: the nearest primary survey behind it polled only people who already use AI tools. The concrete corroboration is Yext's 2026 consumer research, which found 28% of people had tried a new local business in the past six months because an AI tool recommended it. The brands that get cited win that consideration. The brands that don't lose share even when they have better products. But "get cited" is engine-specific work, and the engines are not converging.

What to actually do

Stop thinking about "AI visibility" as a single surface you can measure. Track what each model says about your brand, separately, the way you'd have tracked AltaVista and Yahoo separately in 1998. At best, two engines agree on six citations in ten. At worst, one in six. That’s a big gap that needs filling with engine-specific work: schema and entities for ChatGPT, freshness and community presence for Perplexity, forum and video presence for the AI Overviews etc.

The short version:

• Cross-engine citation overlap runs from 59% down to 16% (BrightEdge, April 2026), and Google's AI Mode shares just 14% of its URLs with Google's own top ten (SE Ranking).
• Each engine rewards different signals. There is no single "AI visibility" lever.
• A blended score can go up while your visibility on the engine that matters goes down.

If you don't know what each model says about you, the free LLM audit runs your brand across GPT, Claude, Gemini and Perplexity and tells you where you show up. Separately, not flattened into a score.

There's a clock on this too. Search Engine Journal made the case this week that the cheap phase of AI citations won't last: OpenAI is already testing ads and building the click-tracking plumbing a paid layer runs on, and the piece is written by someone who watched Google fence off press-release SEO overnight in 2013. Worth remembering how that one ended. The brands already standing on the field when the fence went up kept the best ground. The citations you earn now are that ground.

The single-score dashboards will catch up eventually. Until they do, the brands that track engines individually will be the ones that show up in the answers. The ones that don't will keep watching a number go up while their actual visibility goes sideways.

Make your brand the answer.