What the Wellows citation study changes about measuring AI visibility

#ai #chatgpt #seo #marketing

Sites cited across a topic also tended to be cited on separate questions about it in Wellows’ study of 9,471 questions. That finding describes citation patterns; it does not prove that publishing more pages earns more citations. Study

I think this is the most useful development in this week’s AI search news for a founder deciding where to put a small marketing budget. It gives us a reason to improve how we measure visibility before approving another batch of articles.

The practical question I would ask is simple: are we missing useful content, or are we failing to get cited for questions our content already answers? Those problems deserve different work.

What happened, and when

The Wellows research page lists September 28, 2026 and an October 7 update. Its analysis covers English-language questions collected from January to May 2026 across ChatGPT, Gemini, Perplexity, Google AI Overviews and AI Mode. This week’s publication activity should not be mistaken for a fresh October measurement. Dates and method

Researchers separated questions within each topic into measurement and held-out groups. They checked whether citation coverage on one group was associated with citations on the other. Both groups came from the same period. The sample included websites already cited in the measurement group, so it cannot explain how an uncited website earns its first citation. Method and scope

That last qualification matters to me. If your company has never appeared in the answers you track, I would not use this study as a forecast of what your next content project will achieve.

The distinction I would keep in every report

Wellows measured citation coverage, not published content depth. Outside-topic citation reach had the strongest association with held-out citations on every engine. Among sites grouped by reach, ChatGPT had the weakest coverage relationship. These are observational results, not established ranking factors. Findings

My interpretation is that a content plan and a visibility report need separate jobs.

A content plan should tell me which buyer questions we can answer well. A visibility report should tell me what an assistant actually returned when we asked those questions. I would never accept the existence of a page as proof that the company is visible.

I would also keep being named separate from being cited. For a consultancy helping brands get named by assistants, I want to record whether the answer mentions the company and whether it links to the company’s website. Those are different observations. I would not silently substitute a citation metric for the business outcome a founder asked me to track.

What I think coverage misses

The tempting response to a study about topical authority is to commission a bigger topic cluster. I would first ask what evidence supports that decision for this particular company.

Suppose, as a hypothetical example, a specialist supplier already has a detailed page answering a buyer’s compatibility question. The assistant names another supplier. My first move would be to inspect that answer and its sources. Writing another version of the same explanation would be an untested response to the gap.

The reverse situation calls for different work. If the company cannot point to a clear, accurate answer to the question, I would improve the information before treating the absence as a distribution problem.

The second-order effect I care about is how teams allocate work. If we combine content completeness and observed visibility into one score, I think we make it harder to decide who should act. The content owner needs a specific missing answer. The person investigating citations needs the saved assistant response and the sources it used.

I also want a later test. I would not treat a relationship observed earlier in the year as a current operating rule without checking it against the questions our buyers ask now.

Four steps I would take before expanding the content plan

1. Choose buyer questions and keep them stable

I would start with questions that affect a real buying decision. For example, a founder could ask sales staff which questions repeatedly delay a purchase, then select the ones the company can answer from direct knowledge.

I would save the exact wording and keep the location and collection settings consistent. If I change the question, I would label it as a new test.

The reason is comparability. I want to know whether the observed answer changed. I do not want a rewritten prompt to become an apparent marketing win. I would also retain disappointing answers, rather than replacing them with prompts that make the company look better.

2. Map each question to the best existing answer

For every selected question, I would identify the page that should help a buyer answer it. Then I would read the relevant passage without relying on the page title.

Does it answer the question directly? Can the company support what it says? Where there is a limitation, does the passage explain it?

I would mark a missing or incomplete answer as a content task. If the answer is already useful, I would record that separately from whether an assistant cited it.

The reason is to avoid prescribing more writing before diagnosing the gap. This is also where I would resist making a page merely to satisfy a spreadsheet row. The proposed page should have a clear use for a buyer.

3. Keep an engine-specific record of mentions and citations

I would save each answer with its collection date and the engine used. Beside it, I would record whether the company was named and which source pages supported the response.

I would review each engine separately before producing any combined summary. If the company appears in Perplexity but is absent from ChatGPT, I want that difference visible to the person choosing the next task.

The reason is decision quality. A blended score can make a specific weakness look like an average result. I would rather give a founder a small set of inspectable answers than a dashboard whose headline cannot be traced back to the evidence.

4. Make a bounded change and leave a comparison untouched

I would choose a content change with an explicit reason. Perhaps the existing page leaves a buyer’s question unanswered. I would record the edit and continue collecting the same prompts. I would leave comparable pages unchanged where practical.

Before making the edit, I would define what I hope to observe: a more accurate description of the company, or a citation to the improved answer. I would keep those outcomes separate.

The reason is to make the result reviewable. If visibility improves, I want to examine whether untouched pages changed too. If it does not, I want a clear record of what we tried. Either outcome should help decide the next piece of work without turning a single answer into a promise.