How to read an AI answer like an analyst

Most people read an AI answer the way they read a results page: scan for their own name, feel relieved or annoyed, close the tab. An analyst reads it as a document with a structure. Who asked, what the engine decided, which pages it leaned on, and what it did not say. The whole read takes ten minutes once you have the habit, and it ends with something a colleague can act on.

We will use the same example the rest of the site uses. It is invented, but it is shaped like the real thing.

For alpine climbing, the most recommended hardshells are the Alpsteel Ridgeline AR for durability, the Northlark Solano as a value pick, and the Ridgewell Cirrus Pro as a lighter option. Reviewers at GearBench and TrailNotes rate the Ridgeline AR highest for sustained bad weather, while the Cirrus Pro is praised for mobility on technical routes.

Step 1: read the question, then its intent

Start with the question, not the answer. “What’s the best waterproof hardshell jacket for alpine climbing?” is a buying question near the end of a decision: the person knows the category, has a use case and wants a shortlist. That is transactional intent, and it deserves more of your attention than “how do hardshells work”.

Ask two things. Is this a question our buyer actually asks, in these words, in this market? And how much is it worth? Bluemoon weights questions by buying intent, using public commercial signals as an input; the weight is an estimate and is labelled as one. If the question is low intent or not really yours, note it and move on. Analysts spend their time on the answers that matter.

Step 2: list every brand, with its role

Now read the answer once for names only. Write down every brand, then assign each a role. The role is the part people skip, and it carries most of the meaning.

Two habits help here. First, record absent brands deliberately. If you sell in this category and are not in the answer, that is the most important row in the table. Second, do not upgrade a mention into a recommendation because the adjective was kind. “Praised for mobility” is a mention. “The one to buy” is a recommendation. The distinction is what makes the count over time meaningful.

Step 3: find the sources the answer relies on

Every displayed source is a clue about where the engine gets its opinion. In the example there are three: gearbench.example, trailnotes.example and alpstore.example. Two are editorial reviews and one is a retailer. None of them is a brand’s own site.

For each source, ask who controls it, whether it is cited by other engines for the same question, and whether your brand appears on that page at all. In our example, gearbench.example is displayed by 3 of the 6 engines we track for this question, the other two by 2 of 6. A source cited across engines is a place where a single change can move several answers at once.

Note the sources you do not see as well. If your own product page never appears, the engines are learning about you second-hand. That shapes what kind of brief you write in step 5.

Step 4: measure the gap, across engines

One answer from one engine is an anecdote. The gap becomes a finding when you line the same question up across every engine you track and count roles. Here is the example on one day.

Read it as two gaps, not one. There is a presence gap (named in 3 of 6, absent in 3) and a role gap (named three times, recommended never). They usually need different fixes. Presence is about whether the engines have anything to retrieve about you for this question; role is about whether what they retrieve says you are the right choice.

Step 5: write one brief

The purpose of the read is a single brief with a single owner. Not five ideas. One. There are only three kinds, and the sources from step 3 usually tell you which one you need.

In the example, Ridgewell is not recommended anywhere, and the most shared source is an editorial review it appears in only in passing. That points at a placement: get the Cirrus Pro properly reviewed on gearbench.example, the page that 3 of 6 engines rely on for this question and, in our fictional data, for 14 other tracked questions. Owner: PR. Priority: high, because the intent is transactional. Check: the role count for this question over the next 30 days.

What not to do

• Do not write a brief per answer. Group by question and source; one placement often closes several gaps.
• Do not rebuild a page because one engine ignored it on one day.
• Do not read a mention as an endorsement. Roles exist because the words matter.
• Do not chase every source. A retailer page you cannot influence is a fact, not a task.

The first few answers take a while. After a week the habit becomes fast, and the conversation in your team shifts from “are we in ChatGPT” to “which page do we need to be on”. That is the whole point.