| name | geo-answer-page |
| description | Write the canonical answer page a brand gets cited for — the durable, structured, sourced answer that social posts point into and that AI answer engines can quote. Use when the user wants to WRITE or RESTRUCTURE one specific page so an engine can quote it: 'write the answer page for X,' 'make this page citable,' 'answer block,' 'add a limits section,' 'who reviewed this page,' 'canonical answer,' 'the page our social posts point into,' 'rewrite this guide so ChatGPT can quote it.' Covers the answer-block structure, the limits section, sourcing and reviewer requirements, page-level schema, machine-readable files, and the tension between Google's guidance and what other engines reward. Page-writing only: for site-wide GEO/AEO/LLMO strategy, 'how do I show up in AI answers,' llms.txt and agent-readiness see upstream ai-seo first; for generic structured data see schema; for topic-cluster planning see content-strategy; to measure whether it worked, upstream ai-seo has the monitoring chapter. |
| metadata | {"version":"1.1.0"} |
GEO Answer Page
In this pack GEO means Generative Engine Optimization — being retrievable and quotable by systems that answer questions — not pretending to be geographically local.
The distinction that drives everything: traditional SEO gets you ranked; GEO gets you cited. A well-structured page can be quoted while ranking on page two, because these systems select passages on quality and extractability rather than on rank position alone.
The stakes, with their scope attached rather than as a blanket percentage — Pew Research (Jul 2025, real browsing data from 900 US adults across 68,879 queries) found people clicked a traditional result on 8% of visits where an AI summary appeared versus 15% where it did not, that only 1% clicked a link inside the summary, and that 26% ended the session entirely (versus 16%). Ahrefs (Feb 2026, 300K keywords, desktop, informational intent) measured −58% CTR at position 1 where an AI Overview appears, tapering to −19% at position 10. The honest summary: on queries where an AI answer appears, clicks fall by roughly half, and the magnitude varies a lot by query type.
This page is also the anchor for every other channel in the pack. Social attention decays in days. A canonical answer compounds, and it is where the attention should land.
Read first
If .agents/brand-context.md exists, read it — the claim ledger, prohibited-language list and
channel CTA policy. If it does not, do not stop: derive what you need from the codebase as you
draft, and run claim-check on the finished draft, which builds the ledger properly.
The honest tension, stated once
Google's own guidance says: no special markup or files are required for AI Overviews; don't chunk content for AI; don't write separate content for AI (that risks the scaled-content-abuse policy); helpful, people-first content wins.
Other engines behave differently. ChatGPT, Claude and Perplexity actively reward extractable structure — self-contained passages, comparison tables, definition blocks, FAQs — and they parse machine-readable files when present.
The resolution is not a compromise, it is a sequence. Write for a person, organise for clarity — and the organisation that helps a hurried human read is the same organisation that makes a passage extractable. Do not write AI-specific variants of anything. If a structural choice only makes sense for a machine, it is the wrong choice.
What earns citation
The Princeton GEO study (Aggarwal et al., KDD 2024) measured nine optimisation methods on GEO-bench, with GPT-3.5-turbo as the generative engine. Relative lift on the paper's headline metric, Position-Adjusted Word Count:
| Method | Lift | How |
|---|
| Add quotations | +41% | Named expert, with title and organisation |
| Add statistics | +31% | Specific numbers, with their source and date |
| Fluency | +28% | Readability and flow |
| Cite sources | +28% | Authoritative references, linked |
| Technical terms | +18% | Correct domain vocabulary |
| Improve clarity | +14% | Simplify without hollowing out |
| Authoritative tone | +10% | Paper's own verdict: "no significant improvement" |
| Unique vocabulary | +6% | Filed under "non-performing" — skip it |
| Keyword stuffing | −8% | Non-performing; −10% on the Perplexity replication |
Four things to take from this. First, quotations and statistics are the strongest single methods — not citations. Cite sources is comparatively weak alone (the paper measures it 8% below quotation addition) but strong in combination, so treat it as a multiplier rather than the first move. Second, fluency plus statistics is the strongest pair, beating any single method. Third, low-ranked pages gain the most: Cite Sources produced a +115% visibility gain for pages ranked fifth in SERP, while top-ranked pages lost around 30% when everyone optimises — the opposite of the usual "you need domain authority first" story. Fourth, keyword stuffing and unique-word padding are the paper's two explicitly non-performing methods; stuffing measured 10% below baseline on the Perplexity run.
Treat these as directional and dated. A separate 200-query replication on Perplexity.ai gave smaller and differently-ordered results (+22% best on the same metric), which is itself the warning: the ranking is engine-specific, it is from 2024, and retrieval behaviour moves.
Structure of the page
1. The standalone answer, first
Open with a 40–60 word block that answers the question completely and works with nothing around it. No preamble, no "in this guide we'll explore". If a machine lifts only this paragraph, it should still be correct and attributable.
Whale eye is when the whites at the side of a dog's eye become visible, usually because the head is held still while the eyes move away. It generally signals that the dog wants the current situation to stop. It is one cue among several and should be read together with the ears, tail and posture.
Then repeat the question's own phrasing naturally at least once in the body — that is the string being matched.
2. Headings that match how people ask
H2s and H3s phrased as queries, not as internal taxonomy. If your product calls it a "cue trace" and searchers call it "body language", the heading says body language.
3. A comparison table
Comparison content is the single most-cited format across AI answers. Anywhere the topic contains a distinction — X vs Y, this state vs that state — a table beats prose, because a table is trivially extractable and prose is not.
4. The limits section
A visible "what this cannot tell you" section, and where relevant an explicit escalation boundary — when the reader should stop reading and go to a professional.
This is not a disclaimer. It is a retrieval asset: it is the part of the page that distinguishes you from the confident, unsourced content in the same category, and answer engines that surface a caveat alongside a claim tend to surface the source that provided it.
5. Sources, reviewer, and dates
- Primary sources placed next to the claim they support, not collected in a footer.
- A named reviewer for anything health-, legal-, or finance-adjacent, visible on the page.
- A visible reviewed/updated date, and a real one. Never emit render time as a freshness date — it makes every page claim to be updated today, which is both false and detectable.
6. Internal links
Into the product's actual entry point, and across to adjacent answers in the same cluster.
Cover the cluster, not the keyword. Google's AI features fan a single query out into several related ones and synthesise across the results, so a page that comprehensively answers a parent topic gets retrieved for variants a narrow page never sees. Before writing, brainstorm the five to ten questions the engine is likely to fan out to, and make sure the site answers them.
Schema
| Content | Schema | Why |
|---|
| Guide / article | Article, BlogPosting | Author, date, topic identity |
| Step-by-step | HowTo | Step extraction |
| Q&A section | FAQPage | Direct question matching |
| Product | Product | Price, features, ratings |
| Comparison | ItemList | Structured comparison |
| Embedded video | VideoObject | Only when genuinely embedded |
| The brand itself | Organization | Entity recognition |
The wording must match across every surface — visible page copy, JSON-LD, Open Graph, and llms.txt. Divergence between them is a correctness bug that is invisible to humans and obvious to a parser. It is also the most common way a claim your team already retired stays live for months.
Machine-readable files — and why llms.txt is not one worth building
robots.txt is the one that matters. Verify the citing crawlers are allowed: GPTBot and ChatGPT-User (OpenAI), OAI-SearchBot, PerplexityBot, ClaudeBot and anthropic-ai, Google-Extended, Bingbot. Blocking them means those engines cannot cite you. Blocking a training-only crawler such as CCBot is a separate and defensible choice.
llms.txt has been comprehensively disconfirmed, and this is the clearest "stop doing this" finding in current GEO practice. Five independent studies, three methodologies, all null:
| Study | Scope | Result |
|---|
| SE Ranking (Nov 2025) | ~300,000 domains | No relationship to AI citation frequency. Removing llms.txt as a variable improved model accuracy — it added noise. |
| Ahrefs (Jun 2026) | 137,210 domains, server + bot logs | 28% publish one; 97% of those files received zero requests in May 2026. AI retrieval bots were 1.1% of the requests that did arrive. 12% came from the GEO industry studying itself. |
| Otterly.AI | 62,100+ AI bot visits over 90 days | ~0.1% touched /llms.txt — a third of what an ordinary content page got. They removed the llms.txt checker from their audit product. |
| Evil Martians (Apr 2026) | CDN analysis, two high-traffic sites | Zero markdown requests from GPTBot, ClaudeBot or PerplexityBot — even for files listed in llms.txt. |
| MaxAEO (Feb–Apr 2026) | 240 matched domain pairs, 8 engines | 11.8% citation rate with vs 11.6% without. Inside the noise band. |
No major provider — OpenAI, Google, Anthropic, Perplexity, Microsoft — has ever documented llms.txt as an input. Google's Gary Illyes said on record that Google does not support it and has no plans to; John Mueller compared it to the deprecated keywords meta tag. The most damning single data point in Ahrefs' set: no AI bot ever requested an llms.txt file that did not exist. Nothing goes looking for it.
The one narrow case where it earns its keep: coding agents and agentic pipelines. In Ahrefs' data, Claude-Code out-fetched every AI retrieval bot combined. Agents fetch it when directed, not speculatively — so an unlinked file will not be found. If your product's users point coding agents at your docs, write one for them. For consumer-app AI-search visibility, it is decoration.
If you already have one, the obligation is not to expand it — it is to stop it lying. An llms.txt claiming six languages for a site that ships one is a false claim on the most machine-readable surface you own. Run claim-check over it; it is published copy.
A related caution from the same research: a crawler identifying itself as prompt-injection-survey/1.0 has been studying llms.txt as a prompt-injection vector, since agents tend to trust ingested content. Auto-generated files from a CMS should be read before they ship.
/pricing.md is a better bet than llms.txt for a product with plans, for a non-retrieval reason: agents increasingly compare products before a human visits, and pricing locked behind JavaScript or a "contact sales" wall gets silently filtered out of that comparison.
The August 2026 retrieval shift, and what it means for a small brand
Something material changed in ChatGPT's retrieval in mid-2026 and it points strongly in favour of the work this skill describes.
Reddit's share of ChatGPT Search citations fell from ~3.8% to under 1% in about four days around 14 August 2026 — an ~86% relative drop on one tracking panel, ~95% on another. The mechanism is the interesting part: on 8 August, Promptwatch logged ChatGPT's use of the site: operator inside its background fan-out queries jumping from 0.37% to 16.8% in a single day — and stable at roughly 16–17% since — with fan-out queries per response nearly doubling, from ~1.08 to ~1.83. Those site-scoped queries target official brand sites, documentation, government portals and regulators — and almost never Reddit.
Reddit was not banned or demoted as a domain. Its share of pages consulted held steady at 25–34% throughout. Retrieval was intact; selection changed. Over the same window, institutional and .org sources rose 73% while forums and UGC fell ~70% and press fell ~82%.
Three caveats to carry with this: OpenAI has never confirmed a retrieval change; both tracking firms call their numbers provisional; and there is a precedent for a measurement artefact — a similar apparent collapse in September 2025 turned out to be Google removing the num=100 parameter and breaking third-party instrumentation, not a behaviour change. This one looks like genuine selection rather than instrumentation, but that history is the right reason to hedge.
It is also engine-specific: Google AI Mode saw a gentler ~30–44% slope, AI Overviews were roughly flat, Perplexity showed no clear drop, and Claude does not cite Reddit at all.
The strategic read for a small consumer app is unusually encouraging. What got promoted is precisely what you control — a well-structured, authoritative first-party site that a site:yourdomain.com fan-out query can answer completely. What got demoted is what you cannot control. This is the opposite of the 2024–2025 "just get on Reddit for AI visibility" playbook, and it means the canonical answer page is the better bet than it was a year ago.
Reddit remains valuable — for demand discovery, for the words people actually use, and for search — but treat AI citation as a reason to write the page, not to write the comment.
Citation is not recommendation
Getting cited means your content was useful to consult. Getting recommended — onto the shortlist — is governed by web-wide consensus: reviews, forums, analysts, press. It is largely independent of your own content, and it is the thing that actually converts.
The ladder: retrieved → cited → mentioned → recommended.
A concrete trap: self-promotional "best " listicles can backfire for an emerging brand. In one 100-query B2B study, 69% of AI Overview citations earned by self-promotional listicles appeared in answers that recommended competitors instead. You paid for the retrieval and someone else got the recommendation.
Which is why third-party presence often matters more than your own domain — brands are reported to be several times more likely to be cited via third-party sources than via their own site. Community participation, review platforms, and genuine coverage feed a layer your own pages cannot reach. Do that authentically; fabricated mentions are both detectable and, per Google's own guidance, against policy.
Content types that get cited
Roughly, by share of citations: comparison articles (~33%) · definitive guides (~15%) · original research and data (~12%) · best-of listicles (~10%) · product pages (~10%) · how-to guides (~8%) · opinion and analysis (~10%).
Underperformers: generic unstructured blog posts, thin product pages, gated content (unreadable, therefore uncitable), undated and unattributed content, and PDF-only material.
If you have any original data at all — even small, even about your own product's behaviour — it is disproportionately valuable, because it is the one thing no competitor can restate.
What not to do
- Write separate content "for AI." Same content, both audiences.
- Chunk pages into fragments. Normal headings and paragraphs.
- Mass-generate thin variants. Scaled content abuse, explicitly.
- Fabricate mentions or spam communities for citations.
- Block the citing crawlers while wanting citations.
- Hide the main content behind JS that doesn't render.
- Ship undated content. Undated loses to dated; recency is weighted heavily.
- Let the JSON-LD drift from the visible copy.
Related skills
- youtube-shorts — a video description is a second, high-authority citation surface for the same answer.
- ai-seo, seo-audit, schema, programmatic-seo (upstream) — deeper platform-by-platform detail, technical audits, and markup implementation.
