Reading AI answers

The four things to read

The method Bay Area AEO publishes names the measures directly: the client sees how often they are mentioned, how they are described, which competitors appear and which sources the AI cites. Those four measures cover almost everything a business needs to know from the monthly test, and they are all observable in the answers themselves. None of them is a vanity metric.

Mentions

Count the share of answers that name the business, separately for each platform and each kind of question. A single overall percentage hides too much. A business named in most direct questions but few discovery questions is known to the assistants and has the next step ahead of it: being recommended. A business named often on Perplexity but rarely on Gemini has a gap in what Google's systems read about it, often the Business Profile. A business named for one service but not another has a content gap for the second.

Track position too. Being named first in an answer, with a sentence of description, is worth more than appearing fifth in a list. Note whether the business is in the main answer or only in a list of “other options”.

Descriptions

Read the words. When an assistant names the business, what does it say? “A family-run Thai restaurant in Temescal known for its khao soi” is a description built from specific sources. “A popular restaurant in Oakland” is a description built from very little. Over months, the descriptions should become more specific and more accurate, and should start to include the things the business wants to be known for.

Compare the words with the business's own pages and its reviews. When a phrase from a service page or a common review theme appears in the answer, the source of that description is clear. When the assistant uses a phrase the business does not like (“a budget option” for a premium service), search for where it came from; it is usually a review, an old article or a listing category.

Factual accuracy

Check every fact the assistant states: hours, address, phone, services, prices, insurance, staff. Each wrong fact is a line item for the cleanup routine in Part 2. Accuracy usually improves fastest of all the measures, because listing fixes reach live answers within weeks. A month where accuracy drops usually means a listing has drifted or a platform has re-imported old data.

Competitors

For each discovery question, note who was named. Over several months, a pattern emerges: the same two or three businesses named in most answers for a service in a neighborhood. Study them. Open their listings, their service pages and the sources the assistants cite for them. The comparison usually shows something concrete: a page that states the condition customers ask about, a set of reviews that mention a specialty, a mention in the neighborhood guide the assistants keep citing.

This is not about imitation. It shows which kinds of evidence the assistants use for this question in this place, so the business can supply its own version.

Sources

The sources the assistants cite are a map of where the answers come from. Group them by type: the business's own pages, competitors' pages, listings (Google, Yelp, Apple, Bing), trade directories, local press and guides, community threads. A business whose own pages never appear among the sources has a content or access problem. One whose listings appear but no third-party sources do needs trusted mentions. One that keeps seeing the same local guide cited for its category has found its next outreach target.

Named for the wrong thing

Sometimes the business is named, but for something it does not want to be known for: a full-service restaurant recommended only for takeout, a remodeler named for small repairs, a clinic named for a service it has stopped offering. That is a description problem with a clear cause. The assistants are repeating what their sources stress. Look at the categories on the listings, the services list, the most-reviewed items and the pages that rank. Shift the emphasis toward the work the business wants, in each of those places, and watch the descriptions follow over the next few months.

Reading by place

For a business serving several cities or neighborhoods, split the results by place. The pattern is often uneven: strong in the home city, weak two towns over. The cause is usually local evidence. The business has reviews, pages and mentions tied to one place and almost none tied to the other. The fix follows from chapter 21: real content about the work in the weaker places, reviews from customers there, and local mentions where they can be earned.

Reading through variation

Answers vary from run to run, so a single month's change in one question means little. Look for patterns: several questions moving together, the same change on more than one platform, a change that holds the next month. The repeats described in chapter 47 help here. So does patience: on Bay Area AEO's timeline, the deeper signals take two to four months to build, and their effect shows as a trend.

Turning findings into work

Every finding should lead to an action, and the actions sort neatly into the parts of this handbook.

Signs that the work is landing

Some changes are especially encouraging because they show the assistants have absorbed new evidence. The description starts to include the business's specialty in its own words. The business appears for a condition question it never matched before, after the condition was added to the website. A source the business worked on (a new directory profile, a local guide) starts to appear among the citations. Accurate facts hold steady through a month in which the business changed its hours. Note these in the report, with the answers that show them.

Reading each platform

Platforms differ in ways that affect how to read them. Perplexity's citations make its sources easy to see. Google's AI Overviews come with the map results underneath, and the two should be read together. ChatGPT and Claude sometimes answer from model knowledge without searching, which can mean an older picture of the business; asking a follow-up that prompts a search often gives a current one. Copilot's answers carry Bing's map card, which shows listing accuracy at a glance. Chapter 5 covers the platforms in more detail.

Share of answers

For a business in a competitive category, the most telling single number is often its share of answers for its most important discovery questions, set beside the shares of its two or three main competitors. That comparison shows movement in the market in a way an absolute count cannot. It also keeps attention on the questions that bring new customers.

Writing up the reading

End each month's reading with a short summary in plain words: what changed, what the answers now say, what the work next month will be. Two or three real answers, quoted, make the summary concrete. That summary becomes the core of the monthly report in chapter 49.

Related chapters47. Running the monthly test · 49. The monthly report · 31. Review language · 34. Local press and guides