In our first note we suggested publishing evidence only you hold, and last week we looked at the searches an AI runs on your buyer’s behalf. New research changes what that evidence has to look like, because a citation can wear off within days.
What an AI can’t write for you
AI now writes about half of the new articles published online, by the count of the AI-detection firm GPTZero, so publishing more no longer sets anyone apart. As we wrote from New York in October, the tools are worth using, but language models produce the plausible average of what they’ve read, and we think an engine has little reason to cite a page that only repeats it. Publishing still pays when a page adds something the engines haven’t read, and the more of the web AI writes, the more your first-party data is worth: what your business sees that nobody else does, and your view of what it means.
Say you sell payroll software and want to be the answer when a finance director asks an AI how to cut payroll errors. You used to write an article for every way of asking, and an AI can now write those in seconds. What it can’t write is the error rate across the payruns you processed last quarter, or your read of which companies get payroll wrong most and why, because nobody else has that data.
How long a citation lasts
Profound, which tracks AI visibility, followed 883,000 cited pages across seven AI engines for a year and found the typical page lost half its share of citations within 11 days of its peak. News and research reports faded fastest. Pages cited by five or more engines lasted longer and were more than five times as likely to keep half their peak share after eight weeks. So a report every engine relies on can last, but one big annual report is a risky plan on its own.

Ramp, the US corporate card company, publishes a new number every month instead. Since April 2025 its Ramp AI Index has tracked the share of US businesses paying for AI tools, drawn from its customers’ card and bill payments, with its method and limits alongside. When we asked ChatGPT and Gemini in November what that share is, both cited Ramp’s latest reading, 55.9% for September, before any other figure.
Where the number has to show up
Vercel, the web hosting company, and the analytics firm MERJ watched AI crawlers across Vercel’s network for several months in 2024 and found that none of the major ones ran JavaScript except Google’s, which renders pages through Googlebot. A number that only appears once a page’s scripts have run, such as a price that loads after the rest of the page, never reaches the crawlers OpenAI and Anthropic run. Put the number in the page’s HTML from the start and mark it up with structured data so the engines know what it measures, then check it by loading the page with JavaScript switched off.

Kevin Indig, who writes the Growth Memo newsletter, told Spotlight, a conference on brand visibility in AI search held in London on 13 October, that each doubt a buyer has gets settled on a particular kind of platform, and that comparisons run through publishers and affiliates, who can only include you if you give them comparable specs and data. Give them your data, because a publisher writing your category’s comparison can’t get it from your competitors, and each one that uses it gives the engines another source to check your site against, which is the test Helena Dillon of Microsoft Advertising described at the same conference.
What to do now
Here’s how you can start:
- List what only your business can measure, from what you process to what you see across your customers. If you’re a service business, most of that belongs to your clients, so publish the pattern and the method, never anything that identifies one of them.
- Pick the one that answers a question your buyers already ask an AI, give it a page of its own with the method and its limits, as Ramp does, and update it on a schedule with your read of what changed, so there’s always a new reading to cite.
- Take it to the publishers and affiliates who write your category’s comparisons, as a data pack they can use as it stands.
- Check whether it’s being picked up. Bing Webmaster Tools shows how often Microsoft’s AI cites your pages at no cost, while paid trackers such as Profound, Peec AI, Otterly and Ahrefs Brand Radar estimate from sampled prompts which sources the main engines cite for your category.

If you’d like help with any of this, our AI visibility audit goes further than these steps, testing your buyers’ questions across ChatGPT, Perplexity, Claude and Gemini and checking whether the engines can see your numbers at all.