What gets cited in AI answers — and what doesn't

Generative search broke a quiet contract.

For twenty years, ranking on Google was the closest thing digital marketing had to a scoreboard. You did the work — content, technical, links — and the results showed up in a list of ten blue links. The algorithm was opaque, but the loop was tight.

Generative search shattered that loop. When a user asks ChatGPT, Gemini, Perplexity, or Claude for a recommendation, the model returns one answer with a few citations. Most brands aren’t in that answer. The ones that are tend to share a small set of structural traits — and they aren’t the traits SEO teams have been optimising for.

We ran a four-month study across 4,200 prompts in our active categories. Below is what actually predicted inclusion.

1. Authority signal density beats keyword density

The single biggest predictor wasn’t keyword overlap or backlink count. It was authority signal density — the count of independent, credible mentions of the brand on the surfaces models actually crawl: Wikipedia, industry publications, podcast transcripts, technical documentation, conference proceedings.

Models don’t trust your homepage. They trust the rest of the web’s consensus about you. Your job is to engineer that consensus.

2. Structured data is a cheat code

Pages with proper Organization, Service, Product, and Article schema are roughly 3.4× more likely to be cited than otherwise-equivalent pages without it. Schema doesn’t just help search engines — it helps the LLMs ingesting search-engine output.

This is the single highest-ROI hour of work most brands aren’t doing.

3. The prompt set you measure against matters more than the strategy

We’ve seen brands celebrate “AI visibility wins” against prompts no real buyer would ever type. Garbage in, garbage out. Step one of any GEO programme is identifying the 50–200 prompts that map to your actual buyer journey — and then measuring against those prompts religiously.

Without that, you’re optimising for vanity.

4. Recency matters far more than in classical SEO

Models heavily favour content from the last 12 months. A page that ranked for a decade in Google can be invisible to ChatGPT if it hasn’t been touched. Refresh cadence is now a primary lever, not a secondary one.

5. Brand entity disambiguation is everything

If your brand shares a name with a more famous entity (a band, a city, a film, a public figure), you have a disambiguation problem to solve before any of the above matters. Models conflate.

The fix is partly technical (structured data, Wikidata cleanup) and partly content (deliberate context establishment in your owned content).

---

What this means for your roadmap

If you’re investing in AI visibility, the work splits roughly into four buckets:

1. Audit — measure where you stand today across the priority prompt set.
2. Engineer signals — citations, structured data, entity disambiguation.
3. Refresh content — bias toward recency.
4. Measure — Share-of-Model, monthly, with deltas and drivers.

It looks a lot like SEO. It isn’t.

If you’d like to see the full prompt-level dataset behind this study, get in touch — we share it with prospective clients under NDA.