Answer engine optimization is not about writing longer content. It’s about writing tighter content.
Kevin Indig and AirOps analyzed 16,851 ChatGPT queries across 353,799 pages in 10 industries earlier this year. The headline finding contradicts a decade of SEO orthodoxy: pages that cover 26-50% of ChatGPT’s fan-out sub-queries get cited more often than pages covering 100%. The “ultimate guide” format loses to focused articles that pick 2-3 related angles and own them.
For B2B SaaS content teams sitting on libraries of 5,000+ word pillar pages, this changes the scoping calculus. Covering everything at moderate depth is now worse than covering one thing at high depth. This post walks through the 6 structural answer engine optimization changes the data supports — and how to audit and restructure existing content to earn AI citations. If you’re looking for the adjacent discipline of GEO, we have a separate breakdown of generative engine optimization that covers the broader AI-search category.
Why the “cover everything” playbook stopped working
Classical SEO logic said: one topic, one mega-page, as many subtopics as possible, highest possible word count. Google’s ranking algorithm rewarded topical depth and authority signals, and the reader could scroll to the section they needed.
ChatGPT doesn’t scroll. It retrieves.
When a user asks a question, ChatGPT’s search tool runs the query (and typically 2+ fan-out sub-queries it generates), pulls back around 10 URLs per sub-search, reads through them, and picks which ones to cite in the final answer. The page is not competing for a reader’s attention span. It’s competing to be selected from a candidate pool.
That changes what a “winning” page looks like. From Indig’s analysis of the 815,000-query dataset:
- Retrieval rank is the single strongest citation predictor. A page in position 0 of ChatGPT’s internal search gets cited 58% of the time. By position 10, that drops to 14%.
- Pages cited in all 3 runs of the same prompt had a median retrieval rank of 2.5. Pages never cited: median rank 13.
- Query match (cosine similarity between the sub-query and the page’s best heading) is the strongest content signal. Pages with 0.90+ heading match get cited 41% of the time. Below 0.50 match: 30%.
- Domain authority predicts nothing. Always-cited pages had lower DA than never-cited pages in the dataset.
- Pages covering 26-50% of fan-out sub-queries get cited more than pages covering 100%.
The through-line: AEO is a precision game, not a coverage game. A 2,000-word page that hits one fan-out sub-query with a near-verbatim heading match will beat a 10,000-word page that covers 20 sub-queries loosely.
The 6 answer engine optimization signals that predict citations
Here’s the framework we use to scope and restructure content for AI visibility. Each point maps directly to a signal from the Indig/AirOps dataset or the follow-on “science of how AI picks its sources” research.
1. One page, one fan-out sub-query
Stop scoping content against head terms. Start scoping against fan-out sub-queries.
How to find them: take your target head query (e.g., “B2B SaaS SEO strategy”), run it through ChatGPT, and watch the sub-queries it generates when it’s building the answer. AirOps’ research found that 89.6% of queries trigger 2+ fan-out searches, and 95% of those sub-queries have zero monthly search volume in traditional keyword tools.
This is where most SaaS content strategies are structurally broken. Teams track the head term in Semrush, see volume, commission a pillar page, and never realize that 32.9% of ChatGPT citations come from pages that only appeared in fan-out SERPs — not the original prompt’s SERP.
The practical move: for every commercial head query you care about, map the fan-out sub-queries, then build a focused page per high-value sub-query. Interlink them in a cluster. You end up with 4-6 focused 1,500-2,200 word posts instead of one 8,000-word pillar, and the cluster as a whole covers more of the citation surface.
2. H2s that match the sub-query near-verbatim
On cosine similarity: the Indig study scored each page’s H2-H4 subheadings against ChatGPT’s fan-out queries using bge-base-en-v1.5 embeddings. Pages with heading matches at 0.90+ similarity were cited 41% of the time. Pages below 0.50 similarity: 30%.
An 11-point citation rate swing from one structural change is enormous.
What this looks like in practice: if ChatGPT’s fan-out query for “B2B SaaS SEO” includes “how do you measure SEO ROI for SaaS,” the matching H2 should be “How Do You Measure SEO ROI for SaaS?” — not “Measurement Framework” or “Our Approach to Attribution.” Word order matters. Question format matters. A matching AirOps follow-up analysis found a 2.2x citation lift from title-to-query alignment alone.
Two rules we now follow for AI content optimization:
- If the page targets a sub-query, the primary H2 restates the sub-query in question format, with the key entities intact.
- Sub-H3s follow the same principle for deeper fan-out queries within the topic.
3. Front-load the answer in the first 30% of the page
Indig’s earlier “ski ramp” analysis of 18,012 ChatGPT citations found 44.2% of all citations come from the first 30% of a page’s text. 31.1% from the middle third. 24.7% from the final third. The pattern holds with a P-value of 0.0 — statistically undeniable at scale.
ChatGPT isn’t skimming. It reads deeply (53% of citations come from the middle of paragraphs, not the first sentence). But it weighs the top of the page most heavily because LLMs are trained predominantly on journalism and academic writing, both of which follow “bottom line up front” structure.
The rewrite pattern:
- The answer to the H2 question appears in the first 2-3 sentences of the section, not after 4 paragraphs of setup.
- The page’s core claim or recommendation appears in the first 150 words, not in a “Key Takeaways” box at the bottom.
- Key entities (tool names, brand names, specific metrics) appear early and get repeated consistently — not renamed or paraphrased in later sections.
4. Entity density in the 20%+ range, with specificity bias
Normal English prose has a proper noun density of 5-8%. Heavily cited content in the dataset had entity density of 20.6% — roughly triple the baseline. This is where LLM SEO starts to diverge meaningfully from classical SEO: entity patterns matter more than keyword patterns.
The follow-up “what AI actually rewards” research tightens this: not all entity types are equal.
- DATE is the most universal positive signal (publish dates, year-stamped claims)
- NUMBER is the second most universal (specific stats, counts, percentages)
- PRICE is a strong negative signal in 5 of 6 verticals (except Finance, where fee and rate data is the actual reference data)
- Knowledge-Graph-verified entities (major brands, Wikipedia-backed institutions) were a negative signal in the dataset. High-cited pages were dense with niche, specific entities — a particular methodology, a precise statistic, a named comparison — many of which have no Wikipedia presence
Practical implication for B2B SaaS: naming specific tools, frameworks, percentages, and dates earns citations. Generic “enterprise-grade” language does not. “We increased MRR by 23% in Q3 2025 using Iterable’s segmentation API” is citation bait. “We significantly improved retention through robust automation” is not.
5. Definitive language over hedged phrasing
Citation winners are 36.2% likely to contain definitive language (“is defined as,” “refers to,” “are three types of”) versus 20.2% for losers.
This is a vector database artifact. When a user asks “what is X,” the model searches for strong “X is Y” vector paths. Sentences structured as definitions form tighter vector matches than sentences hedged with “might,” “could potentially,” “some experts argue.”
The rewrite is simple and boring:
- “It’s important to note that attribution can be complicated” → “Multi-touch attribution assigns conversion credit across every channel a lead touched. Three common models exist: linear, time-decay, and U-shaped.”
- “Many marketers believe that…” → “Marketers typically use…”
- “Some studies suggest…” → name the study, author, year.
6. Rank in Google’s top 10 anyway
The Indig “how AI picks its sources” analysis of 98,000+ citations found that pages ranking #1 in Google get cited by ChatGPT 43.2% of the time. That’s 3.5x higher than pages ranking beyond Google’s top 20. AI Overviews show similar retrieval bias toward top-ranking organic results — we broke that down separately in our guide on how to rank on AI Overview in 2025.
This is the part that should temper the “SEO is dead” crowd. ChatGPT’s retrieval pool overlaps heavily with Google’s top results. Classical SEO signals — backlinks, technical health, crawlability, on-page optimization for the head term — still determine whether you get into the retrieval pool at all. The Indig study’s later work showed that ~30 domains own roughly 67% of citations in any given topic, and those domains are almost always the ones dominating classical search for that vertical.
The answer engine optimization playbook isn’t “pivot away from SEO.” It’s “keep winning in Google AND cover the fan-out surface.” You need both.
How to audit and restructure existing content
If you already have a content library, the highest-leverage work is restructuring existing pillar pages, not commissioning new ones. If you want a systematic audit framework that covers technical SEO plus AI-readiness signals, our AI website audit guide walks through the full 5-pillar approach. For this specific restructuring exercise, here’s the workflow:
Step 1. Pull a list of your top 20 pillar pages (>3,000 words, targeting commercial head terms).
Step 2. For each, run the target head query through ChatGPT three times. Note every fan-out sub-query it generates in the reasoning or in the underlying search calls. You’ll get 5-15 sub-queries per head term.
Step 3. Check your pillar page’s H2s against those sub-queries. A rough heuristic without an embedding model: would a reader say the H2 and the sub-query are asking the same thing with slightly different words? If yes, you’re probably above 0.80 similarity. If the H2 is generic or metaphorical, you’re probably below 0.50.
Step 4. Count matches. If your page has 12 H2s and only 3 of them match a fan-out sub-query, the page is 25% citation-viable and 75% connective tissue.
Step 5. For each non-matching H2, decide: does this section have enough unique material to be its own focused post? If yes, extract it, rewrite the H1 to match the specific fan-out sub-query it should target, front-load the answer, and publish as a separate post. Interlink back to the original pillar.
Note: this is early directional data from our own restructuring work, not a formal study. The Indig/AirOps numbers above are the underlying research — check the original Growth Memo piece if you want the full dataset.
What this means if you’re commissioning new content
Four scoping changes we’ve made:
Default length: 1,500-2,200 words. Not because short is inherently better, but because the format forces one fan-out sub-query per page. 8,000+ word guides are now reserved for lead magnets and backlink-bait resources, not AI citation targets.
One head keyword, one fan-out sub-query, one page. If a sub-topic deserves depth, it gets its own page and joins the cluster. Our breakdown of topic clusters for SaaS content goes deeper on how to structure these so they compound.
H2 discipline. H2s get written last, after mapping the page against the fan-out sub-queries it’s trying to own. We rewrite them to match sub-query phrasing as closely as possible without sounding like a robot.
Entity audits before publishing. Count proper nouns and specific numbers in the first 400 words. Target 15+ named entities. If a section is missing specific tools, dates, or stats, it doesn’t ship until it has them.
Frequently Answer Questions
Q: What is answer engine optimization in simple terms?
Answer engine optimization (AEO) is the practice of structuring web content so that AI search tools — ChatGPT, Perplexity, Google AI Overviews, Claude — select and cite it when generating answers. Where classical SEO optimizes for click-through from a results page, AEO optimizes for inclusion in the generated response itself. The core signals are retrieval rank, heading-to-query match, front-loaded answers, and high entity density.
Q: Does this mean I should delete my existing pillar pages?
No. Pillar pages still earn backlinks and can rank for head terms. What you should do is extract the citation-viable sections into focused sub-pages and interlink. The pillar stays as a hub; the fan-out coverage happens in the spokes.
Q: How do I find fan-out sub-queries without an enterprise tool?
Manually, for now. Take your head query, run it in ChatGPT, watch the reasoning or the underlying search calls. AlsoAsked and AnswerThePublic surface adjacent queries but don’t capture ChatGPT’s actual fan-out behavior. AirOps’ commercial tool does this at scale, as do a few newer AEO platforms.
Q: What word count should I target for AEO content?
1,500-2,200 for most posts. The Indig data doesn’t find a magic word count, but the “focused” pattern correlates with mid-length. Longer-form guides are fine if they’re doing a different job (lead magnet, backlink resource) — just don’t expect them to win AI citations.
Q: Does domain authority still matter?
For getting into the retrieval pool, yes — high-DA sites tend to rank in Google, and Google rankings drive ChatGPT retrieval. But within the retrieval pool, DA didn’t predict citation. Page-level signals (heading match, front-loading, entity density) determined which retrieved pages got cited.
Q: Is answer engine optimization different for B2C vs B2B?
Some signals shift by vertical. Indig’s follow-up research found word count matters most in CRM/SaaS (1.59x citation lift for longer pages). Finance is the opposite (shorter pages win). Price data suppresses citations in 5 of 6 verticals except Finance. The 6 structural principles above hold across verticals; the specific thresholds shift.
nswer engine optimization is not a separate discipline from SEO. It’s an optimization layer on top of SEO that rewards precision over coverage, fan-out sub-queries over head terms, and definitive phrasing over hedged prose. The structural shifts aren’t exotic — most of them are things good B2B writers did before the industry got obsessed with word count.
What’s new is the data telling you that the 8,000-word “ultimate guide” has a worse citation profile than three focused 2,000-word posts covering the same ground. If your content team still scopes pillar pages by word count rather than by fan-out coverage, that’s where AI visibility is leaking.
For an adjacent playbook on winning inside specific AI interfaces, our guide on how to rank on ChatGPT covers the citation mechanics in more detail.
If you’d rather have someone audit your existing content library against these 6 answer engine optimization signals, we run a free SaaS content audit — or book 30 minutes at cal.com/onemetrik/30min.
References
This post is built on two pieces of primary research from Kevin Indig’s Growth Memo newsletter, plus the AirOps dataset behind them:
- Kevin Indig — Shorter, Focused Content Wins in ChatGPT. Analysis of 16,851 queries, 353,799 pages, 10 industries. Source for the fan-out coverage data, retrieval rank findings, heading cosine similarity benchmarks, and domain authority results referenced throughout this post.
- Kevin Indig — The Science of How AI Picks Its Sources. Analysis of 98,000 citation rows across 548,534 retrieved pages. Source for the 43.2% citation rate for Google #1 rankings, the 30-domain concentration stat, and the fan-out SERP citation data.
- Kevin Indig — The Science of How AI Pays Attention. Analysis of 18,012 verified ChatGPT citations. Source for the “ski ramp” front-loading pattern (44.2% of citations from first 30% of text), entity density benchmarks (20.6% in cited content vs. 5-8% baseline), and definitive language findings.