Your website gives AI something useful to verify

A buyer asks for a recommendation. The assistant can search broadly, then check a company website for a specific answer. What does it cost? Does it work with the software the buyer already uses? Is there evidence behind the claim?

That gives your own content a practical job. It does not make the rest of the web irrelevant.

What Aiso's data shows

Aiso's study covered 34,960 GPT and Gemini monitoring observations across 75 anonymized projects, with repeated runs from June to September 2026. These were monitoring runs, not 34,960 independent conversations with buyers.

For GPT, the target brand appeared in 49.0% of answers that cited its own website, compared with 2.8% without an own-site citation. The equivalent Gemini rates were 58.4% and 3.8%. All prompts in this comparison were unbranded. Neither group had a subsequent search query naming the target brand.

A citation and a mention can arise together from the same generation process. These figures are associations, not estimates of the lift a website edit will produce. They do not measure purchases or positive recommendations.

The useful question is what evidence your website supplies when the engine checks it.

A site: search is a check, not a guarantee

A query such as site:yourdomain.com pricing limits the search to your domain. If that query is used, the relevant page needs to exist and contain an answer.

The example is illustrative. Aiso's study does not establish a universal growth rate for site:-restricted searches, or a steady decline in Reddit or Wikipedia use. Search behaviour varies with the request and the engine.

Useful content might be a clear pricing explanation or product documentation. For a location-based business, it might be the opening hours, access arrangements or a concrete description of what is available at that address. Adding more general articles is not automatically the answer.

A citation list is not the whole information path

A third-party page can supply information without becoming a visible citation. A URL appearing among retrieved candidates also does not prove the full page was opened or read.

That is why a fall in visible citations from one platform is not enough to conclude that the platform stopped influencing answers. It is also not a reason to abandon a place where your buyers still spend time.

Training is separate again. We cannot inspect the complete training corpus or its update schedule. A model answering a recent question without displaying a web search does not, on its own, establish that its training was refreshed. Retrieval that is not visible to the user, supplied context and other explanations need to be considered.

What I would work on first

Find the buyer questions you should be able to answer. Check what your current pages actually say. Fix missing facts, unclear claims and examples that do not show the result.

Then expand to relevant third-party sources. Reviews, trade coverage and accurate listings can corroborate your own claims and reach buyers you would otherwise miss. Existing third-party work need not stop while you improve your site.

At Aiso, we use the questions customers actually ask and the searches behind them to define what useful content means for a particular business. The priority is the evidence gap, not a rule that every brand needs more blog posts.

What this means for your next page

Pick a question a buyer would genuinely ask. Read the page you would send them. Can they find the answer, see when it was checked and understand the limits? A price without its conditions, or an integration claim without documentation, leaves avoidable gaps.

Then inspect the sources around that question. A review, directory or independent article can provide evidence your own website cannot. Prioritize the missing information rather than assuming onsite work always deserves the entire budget.

Aiso research and methods

The 34,960-observation study explains the monitoring population and limitations. The submitted manuscript reports the full counts in Table 3. It is observational research, not a controlled test of website edits.

For the distinction between retrieved and cited sources, see Aiso's investigation of ChatGPT's information sources. Our guide to checking whether a brand is in model weights explains why behaviour alone does not reveal a complete training corpus or refresh schedule.

Aiso's guide to query fan-outs covers the searches behind an answer. See how Aiso helps identify the questions and content gaps for your business.

Source: Aiso (getaiso.com). Revised on September 26, 2026 to focus on Aiso's own evidence and clarify the distinction between observed associations and causal effects. This edition does not claim to measure a universal decline in Reddit or Wikipedia use, or a universal growth rate for site:-restricted searches.