# Blueprint: building a page that answer engines will cite This is the method behind [skilltrade.marketing/answer-layer](https://www.skilltrade.marketing/answer-layer/), written down so it can be reused. It is deliberately boring. Most of the value is in the sourcing discipline, not in any trick. By Ron Kagan, August 2026. ron@skilltrade.marketing --- ## The premise Answer engines quote things. To be quoted you need three properties, in this order: 1. **A claim worth quoting** that is specific, surprising and attributable to you. 2. **Evidence a skeptic can reproduce**, ideally from a source they already pay for. 3. **Structure that makes the claim liftable** without the surrounding page. Most "AEO checklists" only cover the third. The third is the cheapest and least differentiating. If your claim is the same claim as everyone else's, schema will not save you. --- ## Step 1: find a claim only you can make The Answer Layer page argues that AI Overviews matter more than featured snippets ever did. That is a common enough opinion. What made it publishable was the specific version: > A featured snippet was one slot with one winner. An AI Overview cites several sources. So the > surface roughly doubled while the number of slots a brand can win went up about twenty-fold. The multiple is the claim. Nobody else had computed it, because it requires pairing two columns most people never notice. **The test:** if a competitor could write your sentence without doing your work, it is not your claim. Rewrite until the sentence requires your data. ## Step 2: get evidence from a source the reader already trusts The page uses Semrush's `resource_rank_history`, which exposes a column pair per SERP feature: | Column | Means | |---|---| | `serp__keywords` | Keywords the domain ranks for **where that feature appears** | | `serp__positions` | Keywords where **the domain is inside the feature** | The second column is brand citation, monthly, back to September 2024, at no extra cost. Almost every AI-visibility vendor sells the same measurement starting from your signup date. Two disciplines to carry from here: - **Name the column on the page.** If a claim depends on a distinction the reader might not know exists, put the distinction in the source line rather than trusting them to infer it. - **Re-pull before publishing.** Doing this caught a live page claiming a brand "held 22,550 featured snippets" when 22,550 was the count of keywords where a snippet *appeared*; the brand held 104. Same collapse, wrong by two orders of magnitude, in the first sentence. ## Step 3: check whether your metric can even see the problem The first draft of the Answer Layer page called SEON a counterexample: its citations fell while the surface grew, so it looked like evidence that a bigger board does not win itself. That reading was wrong, and the correction is the most useful thing in this document. SEON's citations are concentrated in the **informational** layer, and thin at the **decision** layer. It is not failing to be cited. It is being cited at the wrong moment. This matters because the queries that used to be separated by weeks now happen in one conversation. A buyer used to search "what is transaction fraud," then "best fraud prevention platforms" a fortnight later, then "X vs Y pricing" once a shortlist existed: three SERPs, three chances to appear, and a funnel stage you could read off the query. Now it is one thread, minutes long, from problem to preference. So an aggregate citation count records a win while the conversation records a loss. **Segment the prompt set by stage** — problem, category, decision — and track the three separately. A brand at 60% informational and 5% decision has a completely different problem from one at 20% across the board, and a single number reports them as identical. Corollary worth building into any plan: the decision layer used to belong to review sites. It is moving to AI answers and to the community threads those answers cite. If your decision-stage asset is a review-site profile refreshed quarterly, that is the thing being disintermediated. ## Step 3b: keep the counterexample Six domains went into the table. Five support the thesis and one does not: SEON's citations peaked and then fell while the surface grew. Leaving it in is not modesty, it is what makes the other five credible. A table where every row agrees reads as selection. A table with a visible exception reads as measurement. Answer engines and hiring managers are both more likely to quote the second. ## Step 4: separate what you can measure from what you cannot The page draws a hard line between retrieval and training data: - **Retrieval** is query-time, cites its sources, responds within weeks, and can be counted. - **Training data** influence is real but lagged, unobservable per brand, and unverifiable. Saying so costs a confident-sounding claim and buys the reader's trust for every other number on the page. If you cannot measure something, say that, and say what you would measure instead. ## Step 4b: verify every quote against the publisher's own HTML Expert quotes are the fastest way to lose credibility, because a misattributed one is trivially checkable and looks like carelessness about everything else. The rule used here: **pull each quote from the primary source's raw HTML, not from a search result, a summary, or another article quoting it.** Doing that on this page caught three things worth knowing about: - A widely circulated version of Amanda Natividad's zero-click content definition is a **paraphrase**, not what her post says. It appears verbatim in multiple third-party write-ups. - Her title has changed since the 2022 piece was published, so the bio on the live page is not the title she held when she wrote it. Attribute to both. - A "276 clicks per 1,000 searches" statistic circulating in 2026 summaries does not appear in either the SparkToro post or the Search Engine Land write-up it is attributed to. Also: **cite the party who disagrees with you.** This page quotes Google's Head of Search arguing that aggregate click volume is stable, directly against its own thesis, then explains how both can be true. A page that only quotes people selling the same conclusion reads as advocacy. One that handles the strongest counter-argument reads as analysis, and is far harder to dismiss. ## Step 5: structure for lifting Only now does the mechanical part matter. - **One question per H2**, phrased the way a person asks it. - **Answer in the first sentence** under the heading. Do not build to it. - **Self-contained claims.** Every sentence with a number repeats enough context to survive being quoted alone. "Elastic held 114 snippets at its peak" travels; "it held 114" does not. - **A table for anything comparative.** Tables get lifted intact. - **FAQ block built from real fan-out queries**, not invented ones. Pull them from a keyword tool's questions report and keep the volume next to each so the choice is defensible. - **`FAQPage` and `Article` JSON-LD.** Cheap, and it removes ambiguity about what is a question. - **An `llms.txt`** at the site root describing the site for models, in the [llmstxt.org](https://llmstxt.org/) format. ## Step 6: instrument it Publishing is a hypothesis. Measure whether the page gets cited: - Track `serp_ai_overview_positions` monthly for your own domain. Free, and it has history. - Run a fixed prompt set against the assistants' citation APIs. Perplexity returns `citations`; OpenAI and Anthropic return `url_citation` annotations from their web-search tools; Gemini returns `groundingMetadata`. The harness for this is published at [skilltrade.marketing/tools](https://www.skilltrade.marketing/tools/). - **Hash the prompt set.** If the questions change, the trend restarts. Otherwise you are comparing different questions and calling it progress. --- ## The checklist - [ ] The core claim requires data a competitor does not have - [ ] Every number traces to a named source, report and date - [ ] Column or metric definitions appear on the page where they could be misread - [ ] Figures re-pulled immediately before publishing - [ ] Every quote verified against the publisher's own HTML, not a summary - [ ] The strongest counter-argument is quoted and answered - [ ] At least one counterexample survives into the final draft - [ ] Unmeasurable things are labelled unmeasurable - [ ] One question per heading, answered in the first sentence - [ ] Claims are self-contained enough to quote alone - [ ] Comparative material is in a table - [ ] Prompt set segmented by funnel stage, not just tracked in aggregate - [ ] FAQ built from real query data, with volumes recorded - [ ] `FAQPage` and `Article` JSON-LD present - [ ] `llms.txt` published at the site root - [ ] Citation tracking running against a hashed prompt set ## What this blueprint will not do It will not get a thin page cited. Steps 5 and 6 are the last 20% of the work and the first 80% of what gets sold as AEO. If you skip steps 1 through 4, you will have a beautifully structured page that says what everyone else says, and answer engines will keep quoting whoever did the work.