Known by name. Missing when it mattered.

A membership organisation that campaigns on a contested area of public policy has published a great deal: research briefings, a podcast, a news archive, accounts of the cases it has won. I was asked whether all of that is packaged so it lands with someone arriving in a hurry, and how it holds up when the answer comes from an AI tool.

I catalogued and read everything they had published, went through their publishing system, and spoke to key stakeholders. Then I put 26 different questions to the AI tools people now use (ChatGPT, Google AI Overviews, Perplexity, Gemini and Claude), several times each. That gave 287 answers, saved and dated as a baseline.

Asked about the organisation by name, the tools quoted its pages on 94 of 135 answers. Asked who campaigns in this area, with no name given, they put it forward in all 15.

But asked the questions someone in real difficulty asks, about the law or their own situation, again with no name given, its pages were cited in none of 137 answers. That included the issues where they have real authority and a body of published work.

The knowledge was there, but it was arranged like an archive, filed by date. The pages being cited instead were built around the question, and they were beatable: fewer than half answered it fully.

A page each for six overarching issues that consolidate their position and answer the questions people typically ask, with everything they've published on the subject gathered underneath. Rather than describe them, I built working versions from their own data, so the leadership team could look at a real page and argue with it.

“It is extremely insightful, and I can absolutely see the value of what you recommend.”

The build is now underway, and once the pages are live we can test the results against the baseline already gathered.