practice

AEO deliverables clients actually accept

The four AEO deliverables that survive a client meeting: executive summary, prioritized fix list, method page, and before/after evidence. Outlines included.

The AEO deliverables clients accept are the ones they can check: a one-page executive summary, a prioritized fix list with owners, a method page that says how every number was produced, and before/after evidence from a re-crawl. What gets rejected, or quietly ignored, is the dashboard of citation counts, the 400-row export, and the screenshot of one AI answer presented as proof.

The difference isn’t presentation. A deliverable is accepted when the person paying for it can repeat any claim in it to their own boss without being caught out. That rules out anything the client can’t verify, anything that implies a guarantee, and anything that needs you in the room to interpret. This post gives the four documents that pass that test, as outlines you can copy, and the failure modes that sink the others.

It assumes you already scoped the work. Scoping and pricing an AEO and GEO audit covers what goes into the engagement; this post is about what comes out of it. For the practice as a whole, start with Running AEO and GEO as a practice.

Why do clients reject AEO reports?

Clients reject AEO reports that promise an outcome nobody controls, or that hand over data without a decision attached. Both are common because the category’s default artifacts are monitoring dashboards and raw crawl exports.

The promise problem is specific to this work. Whether an assistant cites a page is probabilistic and depends on things outside the site. A report that says “these fixes will get you into ChatGPT” is making a claim the client will test, and when the next run of the same prompt names a competitor, the report loses its credibility along with the claim. What prompt tracking can and can’t tell you explains why a single run can’t carry that weight.

The data problem is older. A full crawl export is evidence, not a deliverable. Handing it over moves the prioritization work to the client, who hired you to do it.

RejectedWhyAccepted instead
”AI visibility score” with no methodCan’t be reproduced or explained upwardLens scores with published weights and a method page
Screenshot of one AI answerNext run may differObserved frequency over stated runs, dated
Full crawl export as the reportPushes prioritization to the clientRanked fix list, export as appendix
”Will increase citations”Unfalsifiable until it’s false”Removes these readiness problems; verified by re-crawl”
Generic best-practice checklistNot about their siteFindings tied to their URLs and templates

What goes in the executive summary?

The executive summary is one page with four parts: what was examined, what is blocking, what to do first, and how you’ll prove it was done. If it runs to a second page, the extra material belongs in the findings section.

Outline:

  1. Scope, in one sentence. Domain, number of URLs crawled, scan mode, date. “We crawled 2,140 URLs on example.com on 3 November in quick mode” (an illustrative figure, not a benchmark).
  2. Gates. Anything that stops AI crawlers or search engines from reading important pages: robots.txt blocks on search crawlers, firewall challenges, noindex on money pages, client-side-only content. Each gets one line and the number of affected URLs. If there are none, say “no access gates found” so the reader knows it was checked.
  3. The top five fixes. From the prioritized list, in order, each with the owner type (content, development, or both) and the effort tier.
  4. What will be measured. The re-crawl date and which deltas you’ll report. Name the limits of the measurement in the same paragraph, not in a footnote.

Leave out lens scores on the summary page unless the client already uses them. A number without the method beside it invites the wrong question (“why are we a 61?”). Put scores in the findings section, where the method page sits next to them.

What does a prioritized fix list need?

A prioritized fix list needs one row per fix (not per URL), a priority that someone else could recompute, and an owner. Everything else is optional.

The ranking method this site uses is severity times page value over effort, set out column by column in Prioritizing AEO fixes. Whatever method you use, write its definitions on the first tab of the workbook. A client who disagrees with a row should be able to point at the column they disagree with.

Outline of the workbook:

TabContents
MethodThe scale for each column, the formula, the crawl date and scan mode
Fix listOne row per fix: fix, affected URLs or template, severity, value, effort, priority, owner, status, verified date
AccessOne row per host: each AI crawler you checked, allowed or blocked in robots.txt, reachable or challenged at the edge, intended or accidental
EvidenceThe raw export, filtered to the URLs in the fix list

Three habits make the list survive contact with the client’s team:

  • Group by template. Forty product pages with the same missing markup is one developer ticket. Developers accept tickets that describe a template change; they push back on forty near-identical rows.
  • Write fixes as actions, not observations. “Move the pricing table out of the image into an HTML table” is a ticket. “Pricing content not extractable” is a complaint.
  • Mark deliberate decisions as accepted. If the client blocks a training crawler on purpose, that row moves to “accepted” with the decision-maker’s name. Reporting it as a defect in every monthly report teaches them to skip the access tab. Training vs. retrieval crawlers covers the decision itself.

What evidence proves the work was done?

The evidence that proves the work was done is a re-crawl of the same site in the same scan mode, compared page by page against the baseline, with each fix row marked verified on the date the finding disappeared. That’s the deliverable a monitoring dashboard can’t produce, because it’s tied to specific changes on specific URLs.

Build the before/after section like this:

  1. Baseline. The crawl the audit was based on, dated, kept unchanged.
  2. Re-crawl. Same start URL, same page limit, same path filters, same scan mode. A quick scan compared with a deep scan isn’t a valid comparison.
  3. Summary deltas. Change in each lens score, change in Critical count, pages added or removed. Fewer Criticals is the direction you want.
  4. Per-URL diff for fixed rows. For each row in the fix list marked shipped, the URL’s old and new scores and whether the finding is gone.
  5. Regressions. Every URL that got worse, with a one-line explanation. Clients trust a report that shows its own regressions.

Crawl, fix, re-crawl covers the method in detail, including why a score moves in proportion to the signal’s weight and not the effort spent.

Separate this section from anything about traffic or citations. Before/after crawl evidence shows the page changed. It doesn’t show that the change caused a visit or an answer, and a client who reads it as causal will hold you to the causation later.

How do you report AI visibility without overclaiming?

Report AI visibility as observed frequency over a stated sample, in its own section, labelled as observational. “In 30 runs of 10 questions across two assistants between 1 and 7 November, the client’s domain was cited in 4 runs” (illustrative) can be checked and repeated. “Visibility up 40%” can’t, unless you also give the sample and the variance.

The outline for that section:

  • The question list, fixed for the engagement, with where each question came from.
  • Which assistants, how many runs per question, which dates.
  • Results as counts, not percentages, when the run count is small.
  • A sentence that says what the numbers can’t tell you.
  • Referral sessions from assistants, if the client’s analytics separates them, reported as sessions and not as “AI traffic share” without a denominator.

Measuring AI visibility without overstating it is the longer version of this section. The rule underneath it is short: put observations and fixes in separate sections so neither borrows credibility from the other.

What does the method page say?

The method page states how every number in the report was produced, in enough detail that another practitioner could repeat it. It’s usually the page nobody reads, and it’s the reason the rest of the report holds up when someone does.

Outline:

  1. Crawl settings: start URL, page limit, threads, user-agent, path filters, robots.txt handling, scan mode.
  2. Scoring: what each score measures, with the signal weights or a link to where they’re published. GEOBee’s are on the three lenses page, which can be cited directly.
  3. Severity and priority definitions, copied from the workbook’s method tab.
  4. Visibility sampling: the question list, assistants, run count, dates.
  5. What the audit didn’t cover: logs, off-site mentions, rendering on pages excluded by filters, whatever applies.

The last item matters more than it looks. An audit that names its own limits gets fewer “but did you check…” emails, because the answer is already on the page.

How should the deliverables be packaged?

Package the deliverables as one PDF and one workbook. The PDF holds the executive summary, the findings by template, the before/after section, the visibility section, and the method page, in that order. The workbook holds the fix list and the evidence. The client’s leadership reads the PDF; their developers work from the workbook.

Send the first version as a draft and walk the client through the fix list, not the summary. The fix list is where they’ll find the rows they disagree with, usually on page value, and it’s cheaper to re-rank before sign-off than after. How to audit a site for AEO and GEO covers the audit that produces these inputs.

If you need the evidence tab filled from a crawl you run on your own machine, GEOBee exports every finding with its severity, module, and recommended fix as JSON, and the per-URL scores as CSV. See pricing for the single plan and the 7-day trial.

On this page