# The GEO playbook

Source: https://docs.agentblog.dev/concepts/geo-playbook
Summary: How to write a post an AI search engine will cite, with the strength of the evidence attached to every claim.



Generative engine optimization is writing so that a retrieval system can find
your page, lift a coherent chunk out of it, and attribute that chunk to you. It
overlaps heavily with good SEO, and where it does not, the differences are
specific enough to write down.

Every claim below carries an evidence grade, because the honest state of this
field is that some of it is peer-reviewed and most of it is vendor correlation.
The rank order of the techniques is more reliable than any individual
percentage.

This is the page the `write-blog-post` skill implements. You do not have to read
it to use AgentBlog. It is here because you should be able to check what your
tools are doing on your behalf.

## The one thing that has to be true first [#the-one-thing-that-has-to-be-true-first]

AI crawlers fetch your HTML once and do not execute JavaScript. Vercel and MERJ
instrumented roughly 1.3 billion AI crawler fetches across Vercel's network over
a month and reported that none of the major AI crawlers render JavaScript.
GPTBot requested `.js` files in 11.50% of its requests and ClaudeBot in 23.84%,
and neither executed any of it, which is the detail that makes server logs
misleading: fetching a script is not running it. Googlebot renders. Bingbot
renders on a delay. GPTBot, ClaudeBot, PerplexityBot, and CCBot do not.

**Evidence grade: vendor-reported, from production traffic.** [The rise of the
AI crawler](https://vercel.com/blog/the-rise-of-the-ai-crawler), Vercel and
MERJ, 17 December 2024.

So the first question about any page is not how it is written, it is whether the
text is in the first response at all. Check it the way a crawler does:

```bash
curl -s -A "GPTBot" https://yoursite.com/blog/your-post | grep "a distinctive sentence"
```

If that fails, nothing else on this page matters. Use view source, not the
element inspector, which shows you the page after JavaScript has run.

Four things break this regardless of framework: client-side data fetching in an
effect, content mounted on interaction such as accordions and tabs that render
their children only on click, infinite scroll, and text that exists only inside
an image.

## Structure the page answer-first [#structure-the-page-answer-first]

### The answer capsule [#the-answer-capsule]

Under the H1, and under each H2, write a direct answer of 40 to 60 words
containing no links.

That paragraph is the chunk a retrieval system lifts. A link inside it fragments
the chunk and pulls the reader out of the answer at the exact moment the answer
is being given. Length matters because too short reads as a fragment and too
long stops being liftable.

Answer capsules were present in 72.4% of the blog posts ChatGPT cited in one
analysis of 7,500 ChatGPT referral sessions, and more than nine in ten of those
capsules contained no links. **Evidence grade: industry, correlational.** Adam
Gnuse, [Search Engine
Land](https://searchengineland.com/how-to-get-cited-by-chatgpt-the-content-traits-llms-quote-most-464868),
19 November 2025.

### Question-format headings [#question-format-headings]

Write H2s as the question a person would type. "Do AI crawlers run JavaScript?"
rather than "JavaScript execution".

Across 3 million ChatGPT responses and 30 million citations, cited content was
about twice as likely to contain a question mark, and 78.4% of the citations
tied to questions came from headings rather than body text. **Evidence grade:
industry, correlational.** Kevin Indig's study, [as reported by Search Engine
Land](https://searchengineland.com/chatgpt-citations-content-study-469483), 18
February 2026. The mechanism is more convincing than the correlation: a heading
phrased as a question matches the query shape a retrieval system is matching
against.

Every H2 and H3 needs a stable `id`. That helps readers, enables jump links in
search results, and gives a retrieval system clean section boundaries.

### Section length [#section-length]

150 to 300 words per section, each self-contained. A section that only makes
sense after reading the previous one is a bad chunk, because it will be
retrieved on its own.

## Write for chunk quality [#write-for-chunk-quality]

This is the most important writing rule on the page and the least intuitive.

**Repeat entity names instead of using pronouns.** "AgentBlog writes the sitemap
from content dates" survives being lifted out of context. "It writes the sitemap
from content dates" does not. Human copy editors remove this repetition.
Retrieval systems need it.

The same logic governs the rest of this section:

* Define an acronym in every section that uses it, not once at the top.
* Do not open a section with "as mentioned above".
* Make each section's first sentence say what the section is about, without
  borrowing from its heading.

## Formatting elements, ranked [#formatting-elements-ranked]

Every figure is given against a named metric, because that is where the
secondary literature on this paper falls apart. The GEO paper reports two
metrics, Position-Adjusted Word Count and Subjective Impression, and a summary
that quotes one number without saying which metric it belongs to is why five
blog posts will give you five different rankings.

Both columns are relative to the paper's printed baseline of 19.3 on both
metrics.

| Rank | Element                       | Position-Adjusted Word Count | Subjective Impression | Evidence grade |
| ---- | ----------------------------- | ---------------------------- | --------------------- | -------------- |
| 1    | Quotations from named sources | **+40.9%**                   | +28.0%                | Peer-reviewed  |
| 2    | Statistics with numbers       | +30.6%                       | +22.8%                | Peer-reviewed  |
| 3    | Fluency optimization          | +28.0%                       | +13.5%                | Peer-reviewed  |
| 4    | Outbound citations            | +27.5%                       | +13.5%                | Peer-reviewed  |
| 5    | Technical terms               | +17.6%                       | +10.9%                | Peer-reviewed  |
| 6    | Easy-to-understand language   | +14.0%                       | +6.2%                 | Peer-reviewed  |
| 7    | Authoritative tone alone      | +10.4%                       | +18.7%                | Peer-reviewed  |
| 8    | Unique words                  | +6.2%                        | +5.7%                 | Peer-reviewed  |
| 9    | **Keyword stuffing**          | **-8.3%**                    | +4.7%                 | Peer-reviewed  |

The peer-reviewed figures come from Table 1 of the GEO paper ([Aggarwal et al.,
KDD 2024](https://arxiv.org/abs/2311.09735)), measured on GPT-3.5-turbo against
top-5 Google sources in 2023. Directionally valid, not guarantees on current
engines. Treat the rank order as the finding and the decimals as decoration.

Three rows are worth reading twice:

* **Authoritative tone is weak on the primary metric**, ranking 7th of 9 there.
  Its stronger second number is why it gets quoted as a headline technique. The
  paper's own conclusion is that a more persuasive and authoritative tone
  produced no significant improvement. Writing more confidently is not a
  strategy. Adding evidence is.
* **Outbound citations sit below fluency optimization**, not in a 30 to 40 per
  cent band. The paper's prose groups citing sources with the top methods, which
  is where the inflated number comes from.
* **Keyword stuffing is the only technique measured below baseline.** It is not
  neutral. It measures worse than doing nothing.

Two non-peer-reviewed rows belong alongside these, kept separate because the
evidence is a different kind:

| Element                        | Reported effect                                                             | Evidence grade                                                                                                         |
| ------------------------------ | --------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------- |
| Structural formatting          | 17.3% relative improvement in citation rate across six engines              | Preprint, [arXiv:2603.29979](https://arxiv.org/abs/2603.29979)                                                         |
| Tables present on a cited page | On 40.0% of top-decile cited pages against 28.2% of bottom-half cited pages | Industry, correlational, [Trakkr Research](https://trakkr.ai/trakkr-research/anatomy-of-an-ai-citation), 16 April 2026 |

### Tables [#tables]

Put comparison and specification data in a real table. Not a bulleted list that
describes a table, and not a paragraph that enumerates one. A table is
structurally unambiguous about which value belongs to which row, which is
exactly what a retrieval system needs and exactly what prose obscures.

### Statistics and quotations [#statistics-and-quotations]

One real statistic and one quotation from a named source per post, each cited,
is the floor. Both are the highest-effect techniques in the table, and both are
the two things a language model will happily invent if you let it.

The `write-blog-post` skill carries the instruction never to fabricate either,
in its always-loaded text. Treat it the same way when you are writing by hand.

## Word count [#word-count]

Longer is not better, and the widely quoted claim that long-form posts are cited
three times more often has no source anyone here could retrieve. The primary
data points the other way: across 174,048 pages appearing in 560,346 AI
Overviews, 53.4% of cited pages were under 1,000 words, the average cited page
ran 1,282 words, and the correlation between word count and being cited was
0.04, which is no relationship at all. **Evidence grade: industry,
correlational.**
[Ahrefs](https://ahrefs.com/blog/short-vs-long-content-in-ai-overviews/), 3
December 2025.

Length is a proxy for completeness, not a target. Write exactly as long as it
takes to fully answer the query and its sub-questions, then stop. A 900-word
post that fully answers a narrow question beats a padded 2,500-word one, and the
padding actively degrades the chunk quality of every section it touches.

## Freshness [#freshness]

Recency is worth acting on, and the specific figures in circulation are not
worth quoting. The most repeated one, that 76.4% of ChatGPT's most-cited pages
were updated within 30 days, traces to a study whose publisher's domain no
longer resolves. **Evidence grade: none we could retrieve.** The
`refresh-blog-post` skill is built on the mechanism rather than the number.

Update your best posts on a schedule, and change `dateModified` only when the
content actually changed. A date that moves on every deploy teaches every
consumer of that field to ignore it, which costs you the signal you were trying
to send.

## Internal links [#internal-links]

Five to fifteen contextual links per post. Every new post should also receive at
least one inbound link from an existing post, added in the same commit. Those
numbers are a working range rather than a measured optimum: the figure attached
to them online comes from a vendor page that describes its own data as
synthesized rather than measured.

Orphan posts are the most common structural failure on a blog that otherwise
does everything right. A post with no inbound internal links is discovered only
through the sitemap, and it inherits none of the topical context that makes the
rest of your cluster legible.

Never put a link inside an answer capsule.

## Entity building [#entity-building]

An AI answer engine has to resolve "who published this" to something. That
resolution is what turns a citation into attribution.

* `Organization.sameAs` pointing at profiles a third party can verify: LinkedIn,
  Crunchbase, GitHub, YouTube, Wikidata.
* An author page emitting `Person`, linked from the byline.
* An editorial policy page with a real corrections process.
* Consistent naming. The same organisation name everywhere, spelled the same
  way.

Off-site mentions carry more weight than raw backlink counts for AI citation
specifically. Across 75,000 brands, Ahrefs measured Spearman correlations of
0.664 for branded web mentions and 0.218 for backlinks against AI Overview brand
visibility, with Domain Rating at 0.326. A separate Ahrefs study of the same
brand set put YouTube mentions at 0.737 against ChatGPT brand visibility.
**Evidence grade: industry, correlational, and correlation is not causation
here.** [AI Overview brand visibility
factors](https://ahrefs.com/blog/ai-overview-brand-correlation/), 26 May 2025,
and [AI brand visibility
correlations](https://ahrefs.com/blog/ai-brand-visibility-correlations), 12
December 2025.

Being discussed on Reddit, in a comparison post, or in documentation someone
else maintains is closer to the mechanism than acquiring a link.

## What to avoid [#what-to-avoid]

* **Keyword stuffing.** Measurably worse than baseline.
* **Content behind interaction.** Accordions and tabs that mount their children
  on click.
* **Text only in images.** A crawler that does not run OCR sees nothing. Mirror
  it in HTML.
* **Marked-up facts that are not on the page.** The most enforced structured
  data policy, and violations can earn a manual action that removes rich result
  eligibility entirely.
* **`llms.txt`.** See [why the blog does not ship
  one](/concepts/why-no-llms-txt).
* **Padding.** Every filler paragraph makes the chunks around it worse.

## The pre-publish gate [#the-pre-publish-gate]

Everything above is checkable. See [the checklist](/guides/pre-publish-checklist),
which is what `agentblog audit` runs.

## Caveats [#caveats]

The AI crawler situation changes every quarter. OAI-SearchBot appeared in 2024,
Claude-SearchBot split off in 2025, Cloudflare changed its defaults in September 2026. Any specific list of user agents has a half-life measured in months, which
is why the blog vendors the list and checks it against upstream on a schedule
rather than transcribing it once.

The correlational figures come from vendors, several of whom have a product to
sell. Ronald Sielinski's [Quantifying Uncertainty in AI
Visibility](https://arxiv.org/abs/2603.08924) (9 March 2026) sampled Perplexity
Search, OpenAI SearchGPT, and Google Gemini repeatedly and found that confidence
intervals on a frequently cited domain's citation share span 3 to 6 percentage
points, and that overlapping intervals are the norm for domains appearing to
differ by less than 5 to 7 points. So an intervention that looks like it moved
your citation share by three points may have moved nothing. Treat the
correlational rows as hypotheses worth testing on your own content, not as
settled numbers.

Several figures that circulate widely are absent from this page on purpose,
because their sources could not be retrieved. The full list, with what happened
to each, ships with the `write-blog-post` skill.

***

*This page is licensed [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
Reproduce it, translate it, quote it. Credit `agentblog.dev` with a link.*