How to read an AI citation study before changing your SEO strategy

Start with the research scope

Bill Widmer’s citation study published by Orbit Media reports 13,184 citations and 1,765 answers from 72 prompts across three B2B brands. Its data runs through 23 August 2026. The methodology also explains that Claude used an agent-with-search setup, while the other tools were accessed through APIs, and that some Gemini records identified a domain rather than an exact page.

Those disclosures make the study easier to interpret. They also show why a large citation count cannot, by itself, establish what works for every industry or consumer interface. The study appeared in the digest I reviewed on 26 October; that does not make its observations October data.

Before reading a report’s recommendations, write down its scope in one sentence:

This study observed [defined outcome] for [question set and audience], using [interfaces and settings], during [collection dates].

If you cannot complete that sentence, look for the methodology or ask the publisher for clarification.

What does the percentage count?

A percentage needs a denominator. These measurements answer different questions:

Illustrative example, not client data: two tools each produce ten answers. One returns 100 source links and the other returns ten. Pooling their links gives the first tool roughly 91% of the weight, even though both contributed the same number of answers.

Neither calculation is inherently wrong. The problem starts when a pooled link share is presented as though it measures how often a business appears across tools.

Ask for the numerator, denominator, duplicate-handling rule and treatment of failed or empty responses. Save those definitions beside the metric.

Do the tested questions resemble your customers’ questions?

Read the prompts before accepting a recommendation.

A question naming a brand tests something different from a question asking for suppliers in a category. A comparison question tests something different from troubleshooting a product. Combining them may hide the distinction your business needs to understand.

For your own review, group relevant questions by the decision they support:

• Understanding: what is the problem, and which approaches address it?
• Evaluating: what trade-offs should a buyer compare?
• Choosing: which provider meets the stated requirements?
• Using: how does a customer implement or troubleshoot the solution?

Use customer conversations, support requests and search data to develop these groups. Treat any prompts generated without that evidence as provisional.

A study focused on B2B supplier selection might suggest useful tests for a consultancy. It offers a weaker basis for changing a recipe publisher’s content plan. Relevance matters more than the headline sample size.

Check the interface, settings and collection dates

Record the product, interface, model where available, search setting, language, location and collection date. Mark unknown settings as unknown.

An API test documents the API conditions used. Before applying its findings to customers using a browser application, check whether that application was tested too. Do not assume the two experiences are interchangeable.

For change claims, compare the collection methods as well as the results. If the prompt set or search configuration changed between runs, the difference has more than one possible explanation.

Keep the original responses or permitted exports so a later reviewer can see what was counted. A dashboard percentage alone makes disagreements difficult to investigate.

Separate an observed pattern from a proposed cause

Suppose a report finds that comparison pages frequently receive citations. That observation does not show that adding a comparison table causes a page to be cited.

The sampled questions may have asked for comparisons. The cited pages may also have stronger evidence, better relevance or other characteristics the analysis did not isolate.

Translate the finding into a testable editorial question:

Do our prospective buyers need a comparison that our existing page does not explain clearly?

If yes, improve that explanation using accurate, supportable information. If no, adding a table solely to imitate cited pages creates work without a clear reader benefit.

The content audit guide explains how to turn identified gaps into useful page improvements and record the evidence behind each decision.

Turn the study into one bounded decision

Use this short decision record before assigning work:

1. Observation: What did the research actually measure?
2. Relevance: Which of our audiences or buying questions does it resemble?
3. Uncertainty: What settings, populations or explanations remain untested?
4. Action: What limited change would help the reader even if citations stayed unchanged?
5. Review: What will we compare, and when will we revisit the decision?

For an initial check, keep your questions and collection conditions consistent across repeated observations. Record how often the page appears, which page is cited and what the answer says. Choose a review window that fits the decision; a handful of observations should remain labelled exploratory.

Measure business outcomes separately. A source link, a referral visit and a qualified enquiry are different events. Use the AI referral traffic diagnosis guide when the question is whether visits changed.

Keep Google’s requirements separate from study-derived advice

For Google AI Overviews and AI Mode, Google says a supporting page must be indexed and eligible for a search snippet. Its documentation says there is no special schema or additional technical requirement for those features. Structured data should match the visible content. These are Google’s published requirements and guidance, with a defined product scope.

A third-party citation study can suggest additional questions to investigate. It should not be presented as an official eligibility checklist.

What should you do with the next citation report?

Write its scope sentence, check its denominator and identify one finding relevant to a real customer question. Then decide whether it warrants a small test, further evidence or no action.

That gives the report a practical role: helping you make a better decision about your own site.

Need help deciding which findings apply to your business? Book a 30-minute discussion. Bring one report and the pages you are considering changing, so we can discuss a focused next step.