Search & AI Visibility
Buyers now ask Google, ChatGPT, Gemini and Claude who they should work with. We measure whether you are named, work out why you are not, and do the retrieval, relevance and earned-evidence work the research shows actually moves it, not the folklore that doesn't.
The promise
Evidence first. We publish our workings.
We read 350 primary sources on what actually moves organic rankings and AI citation before we wrote a word of our own service pages, and found that most of the industry's standard advice doesn't hold up under its own evidence. We run the corrected version on ourselves, in public, and on clients.
What it buys you
Server-rendered pages the engines can actually read, honest sitemaps and crawler access, and a site architecture that doesn't dilute its own topical authority. None of the non-Google AI crawlers execute JavaScript; if the evidence isn't in the initial HTML, it doesn't exist to them.
Around 84 percent of non-Google AI citations go to earned media, not owned pages. We place the directory listings, reviews and trade coverage the engines are already trained to trust, rather than trying to out-write them on your own site alone.
Heading-to-query match and a sourced, self-contained answer in the first third of the page are the strongest reproducible signals in the research. Padded FAQ blocks and keyword-stuffed headings measure slightly negative; we don't ship either.
A frozen prompt set run across every engine, monthly, reported as an appearance rate with an interval, not a single lucky screenshot. Day to day, most of an AI answer's sources change entirely, so a one-off check is noise dressed up as a result.
The definition
AI visibility is how often, and how favourably, a brand is named when a buyer asks ChatGPT, Gemini, Claude or Perplexity for a recommendation. It is measured by appearance rate on buyer-intent prompts and the sources the engines cite to justify the answer. Google says its own AI features are "rooted in our core Search ranking and quality systems", not a separate discipline with its own tricks, and its May 2026 guidance is explicit that structured data is not required and that keyword-form rewrites and extra chunking files are unnecessary.
Most of what gets sold as generative engine optimisation does not survive contact with a controlled test. The peer-reviewed record finds citation-oriented rewrites are mostly null in realistic pipelines and can reduce a page's chance of being retrieved at all. Schema markup measured negative for AI Overview citation in a matched difference-in-differences test. What holds up, across the academic and industry data both: close relevance between a page and the literal question asked, a self-contained sourced answer near the top of the page, and presence in the earned, independent sources the engine already trusts.
We set out the full graded evidence, source by source, in Retrieved, Ranked, Cited, including where the evidence disagrees with itself and where a widely repeated claim turned out to be folklore.
Six research streams, roughly 350 primary-source fetches and a live Search Console, Ahrefs and SERP baseline, graded A to U by evidence quality rather than asserted as fact. It sets out what changed between 2024 and 2026, what actually moves organic rankings, what actually moves AI citation, and what we found when we turned the same method on our own website.
Who it is for
FAQs
The evidence says no, or worse. Ahrefs' own controlled test found pages that added JSON-LD schema saw AI Overview citation fall by 4.6 percent; Google's own guidance says structured data is not required for generative AI search. FAQ blocks measured slightly negative for citation in the same industry research, and FAQ rich results themselves were retired in May 2026. We don't sell either as a tactic.
Yes. Most of what AI engines cite they still read via the open web, so a fast, crawlable, well-structured site with genuinely useful content is the foundation both surfaces are built on. The traffic mix is shifting toward AI answers on informational queries specifically, not away from search as a whole.
In practice, yes: three names for the same underlying question, whether a language model reads, trusts and quotes your content. We used to run them as separate pages; we don't any more, because the practice is one discipline and splitting it three ways just repeated the same explanation with different keywords in the title.
We run a fixed set of buyer-intent prompts across every major engine on a set schedule, logged out, and report an appearance rate with an interval, never a single-run screenshot. We also run the full programme on Bridgehead itself and publish the result, including the parts that didn't work, as the case study.
Keep exploring
Speak with us
Thirty minutes with a partner. We'll tell you whether we can help, what the work would look like, and what it's likely to cost, before either of us commits to anything.
