This Memo was sent to 25,823 subscribers. Welcome to +224 new readers! Subscribe to get the free memo weekly or upgrade to Premium for the full archive, research, frameworks, and templates.
In The science of how AI pays attention, I analyzed 1.2 million ChatGPT responses to understand exactly how AI reads a page. In The science of how AI picks its sources, I analyzed 98,000 citation rows to understand which pages make it into the reading pool at all.
This is Part 3.
Where Part 1 told you where on a page AI looks, and Part 2 told you which pages AI routinely considers, this one tells you what AI actually rewards inside the content it reads.
The data clarifies:
- Most AI SEO writing advice doesn’t hold at scale. There is no universal “write like this to get cited” formula – the signals that lift one industry’s citation rates can actively hurt another..
- The entity types that predict citation are not the ones being targeted. DATE and NUMBER are universal positives. PRICE suppresses citation in 5 of 6 verticals and KG-verified entities are a negative signal.
- The one writing signal that holds across all 7 verticals: Declarative language in your intro, +14% aggregate lift.
- Heading structure is binary. Commit to the right number for your vertical or use none. 3-4 headings is worse than zero in every vertical.
- Corporate content dominates. Reddit doesn’t. AI citation behavior does not mirror what happened to organic search in 2023-2024.
Premium subscribers get a SEO content audit tool that scores any page against all 3 parts of this series and generates a custom rewrite prompt.
LinkedIn is an AI citation machine. Take advantage of it.
LinkedIn ranks #2 for AI citations across ChatGPT Search, Perplexity, and Google AI Mode, appearing in 11% of AI responses. With semantic similarity scores of 0.57–0.60, AI models are closely mirroring your LinkedIn content in their answers. Are you taking advantage?
Enterprise AI Optimization lets you track and steer how your brand shows up in AI search, across your markets and domains. See what Semrush sees and stay ahead.
1/ Specific writing signals influence citation, while others harm it.
While The science of how AI pays attention covers parts of the page and types of writing that influences ChatGPT visibility, I wanted to understand which writing-level signals – word count, structure, language style – predict higher AI citation rates across verticals.
Approach
- I compared high-cited pages (3+ unique prompt citations) vs low-cited across 7 writing metrics: word count, definitive language, hedging, list items, named entity density, and intro-specific signals.
- I analyzed the first 1000 words for list item count, named entity density, intro definitive language token density, and intro number count.
Results: Across all verticals, definitive phrasing and including relevant entities matter. But most signals are flat.
What the industry patterns showed:
When splitting the data up by vertical, we suddenly see preferences:
- Total word count was strongest in CRM/SaaS (1.59x).
- Finance was an anomaly with word count: Shorter pages win (0.86x word count).
- Definitive phrases in the first 1K characters was positive for most verticals.
- Education is a signal void. Writing style explains almost nothing about citation likelihood there.
Top takeaways:
1/ There is no universal ‘write like this to get cited’ formula. For example, the signals that lift CRM/SaaS citation rates actively hurt Finance. Instead, match content format to vertical norms.
2/ The one universal rule: open with a direct declarative statement. Not a question, not context-setting, not preamble. The form is “[X] is [Y]” or “[X] does [Z].” This is the only writing instruction that holds regardless of vertical, content type, or length.
3/ LLMs “penalize” hedging in your intro. “This may help teams understand” performs worse than “Teams that do X see Y.” Remove qualifiers from your opening paragraph before any other optimization.
2/ The entity types that predict citation are not the ones being targeted.
Most AEO advice focuses on named entities as a category: pack in more known brand names, tool names, numbers. The cross-vertical entity type analysis below tells a more specific (and more useful) story.
Approach
- Ran Google’s Natural Language API on the first 1,000 characters (about 200-250 words) of each unique URL.
- Computed lift per entity type: % of high-cited pages with that type / % of low-cited pages.
- Analyzed 5,000 pages across 7 verticals.
* A quick note on terminology: Google NLP classifies software products, apps, and SaaS tools as CONSUMER_GOOD, a legacy label from when the API was built for physical retail. Throughout this analysis, CONSUMER_GOOD means software/product entities.
Results: DATE and NUMBER are the most universal positive signals. Interestingly, PRICE is the strongest universal negative.
What the industry patterns showed:
- DATE is the most universal positive signal, with the exception of Finance (0.65x).
- NUMBER is the second most universal. Specific counts, metrics, and statistics in the intro consistently predict higher citation rates. Finance (0.98x) and Product Analytics (1.10x) mark the floor and ceiling of that range.
- PRICE is the strongest universal negative. Pages that open with pricing signal commercial intent. Finance is the sole exception at 1.16x, likely because price here means fee percentages and rate comparisons, which are the actual reference data financial queries are looking for.
- CONSUMER_GOOD (software/product entities) is mixed. In Healthcare, product entities signal established brands and tools. In Crypto, naming specific protocols and products is core to answering technical queries.
- PHONE_NUMBER is a positive signal in Healthcare (1.41x) and Education (1.40x). In both cases, it is almost certainly a proxy for established brands/institutions/providers with real physical presence, not a literal signal to add phone numbers to your pages.
The Knowledge Graph inversion deserves its own note here:
- The data showed that high-cited pages average 1.42 KG-verified entities vs. 1.75 for low-cited pages (lift: 0.81x).
- Pages built around well-known, KG-verified entities (major brands, institutions, famous people) tend toward generic coverage, which isn’t preferred by ChatGPT.
- High-cited pages are dense with specific, niche entities: a particular methodology, a precise statistic, a named comparison. Many of those niche entities have no KG entries at all. That specificity is what AI reaches for.
Top takeaways
1/ Add the publish date to your pages and aim to use at least one specific number in your content. That combination is the closest thing to a universal AI citation signal this dataset produced. But Finance gets there through price data and location specificity instead.
2/ Avoid opening with pricing in non-Finance verticals. Price-dominant intros correlate with lower citation rates.
3/ KG presence and brand authority do not translate to AI citation advantage. Chasing Wikipedia entries, brand panels, or KG verification is the wrong lever. Specific, niche entities (even ones without KG entries) outperform famous ones.
3/ Heading structure: Commit to one or don’t bother.
We know headings matter for citations from the previous 2 analyses. Next, I wanted to understand whether heading count predicts citation rates and whether the optimal structure varies by vertical.
Approach
- Counted total headings per page (H1+H2+H3) across all cited URLs.
- Grouped pages into 7 heading-count buckets: 0, 1-2, 3-4, 5-9, 10-19, 20-49, 50+.
- Computed high-cited rate (% of URLs that are high-cited) per bucket per vertical.
Results: Including more headings in your content is not universally better. The sweet spot depends on vertical and content type. One finding holds everywhere: Strangely, 3-4 headings are worse than zero.
What the industry patterns showed:
- CRM/SaaS is the only vertical where the 20+ heading lift is confirmed: 12.7% high-cited rate at 20-49 headings vs. a 5.9% baseline. The 50+ bucket reaches 18.2%. Long structured reference pages and comparison guides with one section per tool outperform everything else here.
- Healthcare inverts most sharply. The high-cited rate drops from 15.1% at zero headings to 2.5% at 20-49 headings. A page with 30 H2s on telehealth topics signals optimization intent, not clinical authority.
- Finance peaks at 10-19 headings (29.4% high-cited rate). Structured but not exhaustive: think rate tables, regulatory breakdowns, and advisor comparison pages with moderate heading depth.
- Crypto peaks at 5-9 headings (34.7% high-cited rate). Technical documentation in this vertical tends toward dense prose with moderate navigation structure. Over-structuring breaks up the technical depth.
- Education is flat across all heading counts, which is consistent with the writing signals finding. Heading structure explains almost nothing about citation likelihood in education content.
- The 3-4 heading dead zone holds across every vertical without exception. Partial structure confuses AI navigation without providing the full benefit of a committed hierarchy.
Top takeaways:
Test your vertical’s optimal range from the chart above. If you can’t commit to 5+ headings, use none. A sparse heading structure (1-2 headings) sometimes outperforms zero, but only when those headings divide major sections. Don’t use headings for cosmetic breaks.
