# Blueprint: running a five-step AEO audit

This is the repeatable audit behind
[skilltrade.marketing/answer-layer](https://www.skilltrade.marketing/answer-layer/). It shows
whether AI answers cite a brand, where those citations appear across the funnel, and how much
opportunity there is to grow.

By Ron Kagan, August 2026. Updated September 2026. ron@skilltrade.marketing

---

## Before you start

Choose one domain, one market, and one reporting month. Keep the same definitions throughout the
audit. For assistant testing, write a fixed set of buyer questions and save a hash of the set so a
changed prompt list does not masquerade as progress.

Record every source, field name, database, and pull date. The audit should be reproducible by
someone who did not build it.

## 1. Split presence from citation

Presence and citation answer different questions:

- **Presence:** an AI Overview appeared on a keyword where the domain ranks.
- **Citation:** the domain appeared inside the Overview as a source.

In Semrush's `resource_rank_history`, the distinction is visible in the column names:

| Column | Means |
|---|---|
| `serp_ai_overview_keywords` | Keywords the domain ranks for where an AI Overview appears |
| `serp_ai_overview_positions` | Keywords where the domain is cited inside the AI Overview |

Do not collapse these fields. A report can show Overviews on thousands of a domain's keywords
while the domain is cited in very few of them. Presence sizes the opportunity. Citation shows
whether the brand won a place in the answer.

If another tool uses different labels, verify what each field counts before comparing it.

## 2. Pull your own before and after

Set the domain's peak featured-snippet month beside one current month of AI Overview data. Pull
both the presence field and the citation field for each feature:

| Feature | Presence | Citation |
|---|---|---|
| Featured snippet at its peak | `serp_featured_snippet_keywords` | `serp_featured_snippet_positions` |
| AI Overview in the comparison month | `serp_ai_overview_keywords` | `serp_ai_overview_positions` |

A featured snippet named one source. An AI Overview can name several. Surfer's
[study of 405,576 searches](https://surferseo.com/blog/ai-overviews-study/), updated August 11,
2026, found an average of five sources per Overview. If Overview presence expands while the
brand's citations stay flat, other domains are occupying the added slots.

Across the six anonymized domains on the Answer Layer page, the median increase was about 2x for
keywords carrying the feature and 23x for citations won. Pooling the six domains produced 2x and
15x. Those totals describe this sample, not a universal benchmark.

Keep the awkward rows. One 53x increase started at one featured snippet and one 25x increase
started at two. Both mean "almost nothing to something," not a precise fifty-fold improvement.
Another domain's citation count barely moved while its Overview presence grew. The exception is
where the audit becomes useful.

## 3. Split questions into TOFU, MOFU, and BOFU

A total citation count cannot show where a brand disappears from the buying conversation. Divide
the fixed question set by the searcher's job:

| Funnel stage | Searcher's job | Example question |
|---|---|---|
| **TOFU - problem awareness** | Understand what is happening and why | What is transaction fraud? |
| **MOFU - category education** | Learn which kinds of products solve it | What types of fraud-prevention platforms are available? |
| **BOFU - decision or purchase intent** | Choose a vendor or compare options | Which platform fits this use case? |

One anonymized fraud-prevention platform's AI Overview citations peaked at 544 in June 2025 and
stood at 111 in the August 2026 pull. Many of the remaining citations sat on TOFU questions such
as "what is transaction fraud?" The brand was much thinner at BOFU, where buyers pick a vendor.
One aggregate total hid that difference.

An assistant can carry a buyer from problem to preference in one thread. Measure whether the brand
survives that journey rather than treating every citation as equivalent.

## 4. Measure retrieval-augmented generation

**Retrieval-augmented generation (RAG)** is when an answer system retrieves information from
external sources while generating a response. A linked citation gives observable evidence that
the source surfaced in the retrieval-backed answer. It does not prove that every claim in the
answer is supported by that page.

Run the fixed question set against each provider's API. Each question starts a fresh
request; this does not reproduce a consumer app or a multi-turn conversation.

| Outcome | What to record | What it establishes |
|---|---|---|
| Brand mention | The answer contains a configured brand alias or domain | Visible naming, not necessarily a recommendation |
| Owned-site citation | An answer citation links to the brand's domain or subdomain | A source attached to the answer, not proof of every claim |
| Third-party citation candidate | A cited passage names the brand and links elsewhere | A source to inspect before calling it coverage or support |
| Uncited mention | The answer names the brand and contains no citations at all | Naming without a visible citation; its mechanism is unknown |

These categories can overlap. Preserve the complete answer so mentions without an owned-site
citation can be distinguished from answers without any citation. A search-result URL that
is never cited in the answer earns no citation credit. No mention means no visible naming
in that response; a failed request is unavailable, not a zero.

Training is what a model learned before the question was asked. An uncited mention can
reflect training, uncited retrieval, or another mechanism; it does not prove which one.

The [open harness](https://www.skilltrade.marketing/tools/) preserves each question's ID,
stage, response, sources, provider, model, timestamp and completion status. Keep complete
rates separate from partial observations. Changed questions, aliases, providers, models
or parser settings restart comparisons. Repeated framing experiments interleave matched
questions and disclose missing pairs and variation; they do not establish causation.

## 5. Fix in order

Work from eligibility to evidence to access:

1. **Confirm index and snippet eligibility.** Google's
   [AI features documentation](https://developers.google.com/search/docs/appearance/ai-features)
   says there is no special technical requirement for AI Overviews. A page must be indexed,
   eligible to show a snippet, and follow Search policies.
2. **Publish the specific detail worth citing.** Original data, first-hand examples, precise
   claims, and clear sources give the answer something useful to retrieve.
3. **Stop gating the answer.** Keep the useful core on a public, crawlable page. A download or
   conversation can add depth after the answer has earned the citation.

Then rerun the same monthly fields and the same hashed question set. Measure whether citations and
funnel coverage changed before moving to the next hypothesis.

## What the audit delivers

- A before-and-after view of featured-snippet and AI Overview presence.
- Citation counts kept separate from the size of the answer surface.
- TOFU, MOFU, and BOFU coverage for a fixed buyer-question set.
- Brand mentions, owned-site citations, third-party candidates and uncited mentions reported separately, with completion counts.
- A prioritized fix list: eligibility, specific evidence, then access.

The audit does not compress these into one AEO score. It also does not claim that a citation caused
revenue. Track citation share alongside branded search, direct traffic, self-reported attribution,
and pipeline from the same audience to see whether visibility and demand move together.

## Checklist

- [ ] Domain, market, comparison month, and pull date recorded
- [ ] Presence and citation fields defined separately
- [ ] Featured-snippet peak compared with one current AI Overview month
- [ ] Source fields and calculations preserved for reproduction
- [ ] Private domains anonymized in public outputs
- [ ] Fixed buyer-question set split into TOFU, MOFU, and BOFU
- [ ] Prompt-set hash stored with every assistant run
- [ ] Brand mentions, owned-site citations, third-party candidates and uncited mentions reported separately
- [ ] Failed requests kept unavailable and changed measurement conditions excluded from deltas
- [ ] No claim that an uncited mention proves training exposure
- [ ] Index and snippet eligibility checked before content changes
- [ ] Specific, attributable detail published before structural optimization
- [ ] Useful answer available without a form
- [ ] Same measures rerun after the fix