How to read AI search research

Write down the actual claim

Write down the claim precisely before reading the headline as a business recommendation. Identify whether it concerns access, answers, visits or customer actions.

A headline can drop an important condition, so read the surrounding explanation before deciding what the claim means for you.

Then sort the claim by kind. A measurement reports what someone observed in a stated period. A theory offers a reason for what was observed. A recommendation tells you what to do, and it should rest on one of the other two.

Keep the author's exact words beside your own summary, because a summary drifts each time it is retold. If the claim uses a word such as visibility or authority, find how the author defined it. Two studies can use one word for two different counts.

Find the sample

Look for the question set, businesses and market covered. A result from one industry or country does not automatically describe another.

A large count doesn't remove a narrow sample. If the study doesn't say who was included, mark that missing. Don't assume your business was represented. Ask whether the businesses studied look like yours.

Test the fit on four points. Were the country and language the same as your customers' own? Did the study cover the assistants your customers use? Were the prompts about buying decisions like yours, or about general knowledge? Were the businesses studied a similar size and type to yours?

A no on any point doesn't make the finding wrong. It means the finding says less about your business, and your notes should say so in plain words.

Check when the answers were collected

Find the collection window as well as the publication date. A recently published article can still analyse older observations.

Write the first and last day of collection, not just the publication date. A report published this month may rest on answers gathered much earlier. If the model or product version is given, record that too.

Watch for seasonal effects. Questions about heating, gifts or holidays behave differently across the year. A study run in one season may not describe another, even for the same type of business.

When dates are missing, say so in your summary and treat the finding as undated. Don't fill the gap with the publication date.

Read the denominator

Ask what each percentage divides by. It might count answers, citations, businesses or pages, and those are different populations.

Write the numerator and denominator beside the figure and keep any excluded or failed captures in view. Without them a percentage sounds precise but tells you little.

Some denominators shrink quietly. A claim about how often a type of page was linked may count only the answers that contained a link. Answers with no links then sit outside the figure without comment. Ask what was removed before counting, such as failed requests, duplicates or answers in another language.

An average needs its spread. An average across many prompts can hide a few prompts that behave very differently from the rest. If the study gives no range or breakdown, record that as missing too.

Read charts and examples with care

Check where each chart axis starts. An axis that begins above zero makes a small difference look large. Read the values on the axis before you read the shape of the line.

Look at what each bar or point counts. A chart of totals can rise simply because the study added more prompts over time. A chart of rates needs the same denominator throughout.

Treat screenshots and quoted answers as examples, not as evidence of a pattern. An author can choose the most striking answer from many. Ask how the examples were selected and whether typical answers looked the same.

Be careful with league tables of sources or brands. A list of the most linked sites depends on the prompts asked and the counting rule. Change either one and the order can change.

Check how data was captured

Find out whether the study used consumer experiences, APIs or another collection service. Keep undocumented collection settings as a limit.

Ask whether the full answers and source links were kept. A conclusion from one collection method may not describe another, and missing method details limit how far you can apply it.

Check how many times each prompt was asked. One answer per prompt is a snapshot. Repeated answers to the same prompt show how much the wording and the links move from one request to the next.

Look for whether the study used one account, one device or one location for every capture. That choice is part of the method, even when the author doesn't mention it.

Match the study to the AI experience

A finding about one AI product may not carry over to another. Google says AI Overviews may use different models and techniques, so their responses and links vary.

Google also says AI Overviews are shown only when its systems judge them additive to classic Search, and that they often don't trigger. So a study of AI Overviews depends on which searches produced one at all. Ask how the author handled searches where none was shown.

Location and device matter as well. Google's Search Console help says search results depend on the time, place, device and recent history of the person searching. A study captured from one office on one device describes that setting.

Write the product, the surface and the setting beside each finding. Then compare it only with findings captured the same way.

• Google guidance on AI features
• Google Search Console Help, Performance report overview

Separate change from cause

An observed difference may have several explanations. Look for the comparison design before treating the result as an effect of an intervention.

Look for a control, repeated measurements and other changes in the period, and record which checks the study actually ran. If they can't isolate the effect of the edit, leave the cause open.

Think about what else changed in the period. A new model, a seasonal rise in demand or a change to the prompt set can move a result on its own. A study that compares changed pages with similar pages left alone is stronger than a simple before-and-after comparison.

Be wary of a cause stated with confidence and supported by a handful of examples. Examples can illustrate a pattern. They can't show that the pattern holds across the whole sample.

Revisit the source before any big decision. An older finding can still help, but its age and scope belong beside the claim. Check whether the method was published before you interpret the result.

Grade your summary in plain words. Say whether the claim is well supported, partly supported or untested for a business like yours. Then say what would change your view, such as a published method or a sample from your own country.

Notice who wrote it

Read the author's role and any commercial interest. Those details provide context but do not replace checking the method.

Check whether the article links to its method or underlying evidence. A marketing claim and an independently reproduced finding are different kinds of support.

Ask what the publisher sells. A study that ends by pointing to a product or service has a reason to present its result strongly. That doesn't make the result wrong. It means the method has to carry more of the weight.

Look for a publication date and any later corrections. Research pages are sometimes updated without a note, so save a copy of the version you read.

When two studies disagree

Two careful studies can reach different answers. Before choosing one, compare their samples, dates, definitions and capture methods side by side. The difference often sits in one of those four.

Don't average two findings that used different methods. An average of two unlike numbers describes neither study. Keep both, with their limits, and say which one is closer to your situation.

If neither fits your business, the honest summary is that the question is open for you. That is still a useful result. It tells you where your own small experiment would add the most.

Plan a small experiment

Turn the study into a specific question your own records can address. Keep the experiment small enough to inspect and avoid borrowing its outcome as yours.

Keep the starting record before the owner changes anything, and say what would count as a difference. A study result isn't a promise about your next answer.

Keep the experiment small enough to read in full. One page and a handful of prompts can be reviewed answer by answer. A large set tempts you to read only the summary figure.

A worked reading of one claim

A vendor report says pages with a questions section were linked more often in AI answers. Fern Workshop wonders whether to add one to every service page.

First, the kind. The claim is a measurement with an implied recommendation. Next, the sample. The report covers one country and one assistant, and it lists no dates.

Then the denominator. The figure counts pages that were linked at least once, so pages never linked sit outside it. The capture method is a software interface, with one answer per prompt.

Finally, cause and author. There is no comparison group, and the publisher sells a tool for writing questions sections. Fern Workshop's note calls the claim an interesting association. It is untested for this business, the dates are missing and there is no control.

That note doesn't rule the idea out. Fern Workshop can still add a questions section to one page where customers really ask those questions. It saves the starting answers first and reads the later ones against them.

Sources and method

A Carqo review method. Provider statements link to the relevant documentation.

• Google guidance on AI features Link checked 4 November 2026
• Google Search Console Help, Performance report overview Link checked 4 November 2026

Suggest a correction