The Answer Engine Optimization Playbook
Works with
Implementation kit
- The question list, with five sources and the tags that sort buyers from job seekers
- The citation read and the coverage check, joined in one question ledger
- Seven levers and eight unsupported tactics, each with its source and grade
- The answer page brief, one per question
- The measurement read, with the rule you write before the baseline
- A sample company run from question list to baseline
Buyers ask AI who to hire, and the answer names someone else? List the questions they ask, read who gets cited for each, and measure how often an answer names you.
The full method is below. A question list built from what buyers type, a read of who each engine trusts, a check of what you have already published, levers picked on the evidence, one answer per question, and a read every two weeks on prompts you never edit.
A year of pages published for AI search, and nobody can say whether an answer names the company.
The failure this playbook fixes
How it works
01Pull the questions buyers ask AI
An answer engine names a company in reply to a question, so the unit of work is the question. Pull from at least three sources, such as AI question data with monthly volumes, sales calls, the communities your buyers use, Search Console and your support inbox.
Group the phrasings that are one question in different words, add the volumes, and tag each group as buyer, learner, job seeker, tool shopper or other meaning.
Watch for the highest-volume row. It is often a tool shopper or a job seeker question, and neither turns into a call.
02Read who gets cited, and sort open from owned
Ask each question on at least two engines, three times each, in a fresh session. Log the companies named, the domains cited, the page type and the shape of the answer.
- Open, small sites, individual authors, forums or videos, and the names change between runs
- Contested, a mix of large brands and small sites
- Owned, the same two or three large brands in every run, citing their own pages
- Wrong shape, a definition where you wanted a vendor list, or the reverse
Narrow questions are where a new name gets in. Ahrefs published pages for a brand-new name in 2026. It reached 66.4% of answers to a narrow question and 15.8% to a broader one. Semrush found only 15.2% of 1,094 ChatGPT categories had a clear owner. When a question is owned, add a platform, a company size or a task and read the narrow version.
03Check what you have already published
Search your own live site for each question and grade what comes back as exact, adjacent or none. Record it per surface in one ledger, with the next asset for each row.
We made this mistake ourselves. A question sat at the top of our list marked as unanswered, and an existing playbook and blog section already covered it.
Watch for a ledger that is checked against itself. Read the live site every time.
04Pick three levers the research supports
The kit grades seven levers that reach a small firm. Narrow questions, a plain service page per offer, the third-party lists engines already cite, long-form articles under a person's name, mentions across many sites, specific numbers with sources, and measuring share of appearances.
It also lists eight tactics the research does not support. An llms.txt file showed no correlation with citations across about 300,000 domains (SE Ranking). Adding schema to 1,885 pages produced no lift against matched controls (Ahrefs, 2026).
Almost every large study was run by a company that sells a tracking tool, on brands much larger than a small firm. Treat the levers as the best available bets and measure your own result.
05Write one answer per question
Put the question in a heading in the asker's words, and answer it in the first two sentences. Use specific numbers with their source and date, name the tools and the fields, date the page, and serve it as plain HTML.
Cover what the cited pages leave out. That gap is the reason for an engine to add you.
For questions about who can do the work, the page that gets cited is often the firm's own. For agency questions, service pages were the most cited first-party page type at 30.4% (Ahrefs, 750 prompts).
06Measure the share of answers that name you
Freeze the questions as prompts, word for word, and write down what "it worked" means before the baseline. Run each prompt ten times on each engine in a fresh session, then again every two weeks.
Count a win only when the answer names you. Being cited as a source is useful and is recorded beside it. Semrush found only 21% of the most-cited domains were also the most-mentioned brand.
One answer tells you nothing. Across 2,961 runs of 12 prompts, the same list of brands came back under 1% of the time (SparkToro and Gumshoe).
We run this read on ourselves. Our baseline on October 30, 2026 was zero, with Magnetiz named in none of 799 answers across 27 questions and four engines. A zero baseline still lists the companies each engine names instead.
Share of answers, read on the same prompts every two weeks.
The implementation kit above packages the question list, the citation read, the coverage check, the levers with their sources, the answer page brief, the measurement read and a sample company run from start to baseline.
