The AEO playbook

What the 30-day plan actually does

The plan is four weeks of ordered work for a B-grade brand that wants an A: measure a baseline across all five engines, rank the gaps by how much traffic-weighted query value they represent, rework the source pages the engines should be citing, tighten your schema and entity data, then re-measure against the same baseline. The score is a composite of five factors, and a B grade can come from weakness in one factor or in several.

A grade means you show up early, get described accurately, and get cited when someone asks an AI engine about your category. B grade usually means one of those is broken: you appear but low, or you appear but described wrong, or you are named but never linked. The plan below assumes you already exist in the answers and the job is to raise how you appear, not to break in from zero.

Pondral scores visibility on a published five-factor rubric: Presence (20), Prominence (25), Context (20), Citation Link (20), and Competitive Presence (15). Knowing which factor is dragging your grade down is the whole point of week one, because the fix for a Prominence problem is nothing like the fix for a Citation Link problem.

Week 1: Baseline every engine, then triage

Do not touch a single page until you have a frozen snapshot. Run the same prompt set through ChatGPT, Claude, Gemini, Perplexity, and Grok, and capture three things per query: the verbatim answer, whether you were mentioned and where, and the exact citations the engine pulled. That snapshot is what you measure week four against. Change nothing before you have it.

The engines do not behave the same, so a single blended score can hide your real problem. They differ in how often they link out to sources and in how much they synthesize, so a Citation Link gap can show up on one engine and not another. Read each engine's column separately.

Now triage. Rank gaps by traffic-weighted query value, not by how many gaps you have. One high-intent comparison query buyers actually type carries more weight than a dozen informational long-tails nobody converts on.

1. Build the prompt set from real buyer language: category queries ("best [category] for [use case]"), comparison queries ("[you] vs [competitor]"), and problem queries ("how do I fix [problem]"). Aim for 30 to 60 prompts.
2. Run every prompt through all five engines and log the verbatim response, your mention position, and the citation list for each.
3. Score each result against the five factors so you can see whether you are losing on Presence, Prominence, Context, Citation Link, or Competitive Presence.
4. Weight every gap by the query's commercial value and rank the list. The top five queries are your entire month of work.
5. Save the raw export. Same prompts, same engines, week four. No changes yet.

Week 2: Fix the source pages engines should cite

Take your top five ranked queries. For each one, name the single page on your site that should be the answer. Often that page does not exist, or it exists as a sales page that reads like a brochure. A page that behaves like a reference gives an engine something it can quote. A page that behaves like a pitch mostly gives it adjectives.

The pattern is boring and consistent. Open with a direct, self-contained answer to the query in the first short paragraph, before any preamble, so an engine can lift it whole. Put the comparison in a real table or list, not buried in prose. State facts plainly with dates and numbers the model can quote. A page that answers the question in its first paragraph hands an engine something to lift. A page that warms up for three scrolls makes it dig.

This week targets Prominence and Context. Those factors check where you appear in the response (Prominence) and how you are described (Context). A source page that directly answers the query gives the engine a clear description of you to work from. Whether a given engine picks the page up is something only the week-four re-measure shows.

1. For each top query, create or rewrite one owned page whose sole job is to answer that query.
2. Lead with a short extractable answer, then expand. No throat-clearing above it.
3. Convert any comparison into a scannable table or list. Structure is easier to extract cleanly than prose.
4. Add explicit, checkable facts: prices, dates, specs, limits. Vague claims give an engine nothing specific to quote.
5. Link the page from somewhere crawlable and, if it is genuinely new, resubmit your sitemap so it gets indexed before you re-audit.

Week 2, continued: Schema that machines actually read

Schema does not raise your score by itself. What it does is let a machine parse what your page is, who published it, and what it claims without guessing from the layout. We have not measured its effect on citations.

Add the schema.org types that match the page. Article or the more specific type for editorial and guide content. FAQPage for pages built around question-and-answer blocks, where each Question maps to a real heading. Product with an Offer for anything sold. Organization on your site-wide markup with your legal name, logo, and sameAs links to your verified profiles. Validate every page in Google's Rich Results Test and the Schema.org validator, and fix anything that throws a warning, not just the errors.

1. Mark up editorial and guide pages with Article (or the closest specific subtype) including author, datePublished, and dateModified.
2. Use FAQPage only where the visible page really is questions and answers, with each Question tied to an on-page heading.
3. Add Product plus Offer to anything for sale so pricing and availability parse cleanly.
4. Put Organization on your global markup: legal name, logo, and sameAs pointing to your real external profiles.
5. Run both validators on every changed page and clear warnings as well as errors before moving on.

Week 3: Entity consistency and off-site reinforcement

Engines resolve you to an entity before they decide what to say. If your name, description, and category read differently across your site, your LinkedIn, your directory listings, and Wikidata, the engine has conflicting descriptions to reconcile and can hedge or confuse you with a similarly named competitor. Pick one canonical name, one one-line description, and one category label, and make every surface match it exactly. It is cheap work and easy to skip.

Then reinforce off-site. A claim that appears only on your own domain has one source behind it. The same claim on independent third-party sources has several, and engines that search the web read those pages too.

Auditing your existing third-party mentions is a good use of half a day. Where engines are citing stale or wrong facts about you, the page they cite may be fixable by asking its owner.

1. Write your canonical entity block (name, one-line description, category) and reconcile it across your site, social profiles, and every directory.
2. Create or update your Wikidata item, and seed or clean up Crunchbase, G2, and the two most-cited directories in your category.
3. Pull the citation list from your week-one audit and note every third-party page an engine used to describe you.
4. For any cited page carrying outdated or wrong facts, contact the owner with the correction and the source. Start with the pages engines already cite.
5. Pitch one genuinely useful contribution (data, an interview, an expert answer) to the highest-authority site in your category. Favor one placement on a trusted domain over many low-authority mentions.

Week 4: Re-measure against the frozen baseline

Re-run the exact prompt set from week one through all five engines. Compare gap by gap against the frozen snapshot, factor by factor, so you can see what moved and line it up against the work you did. AI answers vary between runs, so treat small moves (on Pondral's dashboard, changes under 15 points) as directional rather than as proof the work landed. How many gaps close in one cycle depends on how far behind you started and how competitive the category is.

For anything that did not move, read the verbatim response before you conclude the work failed. Common reasons a gap did not close are below, and each one calls for a different change. Do not re-do the same page hoping for a different answer. Diagnose first.

1. Source page exists but reads like marketing, so nothing quotes it: rewrite the opening as a plain, extractable answer.
2. Entity still confused with a competitor: tighten name and description consistency and add the missing sameAs links.
3. You appear but low (Prominence): check whether the query wants a comparison or a list that your page does not provide in a liftable format.
4. You are named but your site is never linked (Citation Link): check whether you have a page that answers that query directly. If not, that is the week-two work.
5. The prompt asks for something you genuinely do not offer: drop it from the target set instead of forcing a page.

Make it a monthly rhythm, not a project

AI visibility drifts because the engines change, competitors keep publishing, and your own facts go stale. A score you measured in April is not a score you can assume in July. Treat the four-week loop as a recurring cadence: a lighter monthly re-audit on the same prompt set, plus a full triage each quarter.

Between cycles, watch the two factors that can move without you touching anything. Competitive Presence can slide when a rival publishes a stronger page for a shared query. Citation Link can drop when an engine stops linking to your pages in its answers. A monthly re-audit is where you see both.