What is generative engine optimization?

Where does the term GEO come from?

The term comes from a 2023 paper, GEO: Generative Engine Optimization, by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, later accepted at KDD 2024. It treated an answer engine as a black box and measured how rewriting a source changed how much of that source showed up in the answer.

The authors tried nine rewrites, from adding quotations and statistics to stuffing keywords, on an answer engine they built from the top five Google results and an OpenAI model, and checked a sample on Perplexity. Adding quotations, adding statistics and citing sources did best: the strongest rewrites raised their visibility measures by about 40%. Keyword stuffing did worse than leaving the page alone. The gains were largest for sources that ranked lower to begin with.

Read that carefully. The measures were the share of words in an answer that came from a source, weighted by position, and a model's impression of how prominent the source was. That is not traffic and it is not a citation count on a live assistant. The rewrites were made by a model, on one test engine, in 2023. "Up to 40%" is a ceiling from a lab, not a promise for your pages.

How is GEO different from SEO?

A search engine ranks pages and sends a person to one. An assistant reads several pages, writes one answer, and may name a few sources under it. The page still has to be fetched and indexed, so everything SEO asks for still holds. What changes is the unit that wins: a passage that answers one question cleanly, not a page that covers a keyword.

Our view, from reading the sites below: most of GEO is SEO that was always worth doing and mostly was not done. The genuinely new part is writing so a machine can lift one answer without the rest of the page.

What do assistants actually need from a page?

They need the words to arrive in the HTML, a page structure that says what each part is, and passages that answer a question without leaning on the paragraph before. On the pages mttr has read, the basics mostly hold and the answer shapes mostly do not: 89% of key pages carry their words in the HTML, but only 9.9% of article pages put a short answer right under a question heading.

The weakest are the cheapest to fix. A heading that asks the reader's question and a paragraph that answers it is a writing habit, not a rebuild. So is a table: only 9.8% of article pages have one. A score or a rule passed says pages are well formed. It does not say an assistant cites the site.

<h2>Our roof repair services</h2>
<p>At Acme Roofing we have been proud
to serve homeowners across the area for
many years, and we believe every
customer deserves honest work. Roofs
come in many shapes and sizes, and so
do the problems they face, from storms
to age to simple wear. That is why we
always recommend calling us first, so
one of our friendly team can come out
and take a look.</p>

<h2>What does a roof repair cost?</h2>
<p>Most roof repairs cost far less than
a new roof. A small leak or a few lost
shingles is usually fixed in a day;
storm damage across a whole slope can
take longer and cost more. We quote
every job for free after we have seen
the roof.</p>
<table>
  <tr><th>Job</th><th>Usual time</th></tr>
  <tr><td>Small leak</td>
      <td>Half a day</td></tr>
  <tr><td>Storm damage</td>
      <td>One to two days</td></tr>
</table>
<p>Updated <time datetime="2026-10-01">
October 1, 2026</time>. Source: <a
href="https://www.nrca.net/">National
Roofing Contractors Association</a>.</p>

Which lines of work are easiest for an assistant to read?

The understand stage of mttr's score asks whether a machine can tell what a page says and lift an answer from it. Across the 287,847 ranked sites it has a median of 60 out of 100. Clinics and dentists lead at 71; News and publishers trail at 56.

See the understand stage in depth for what it holds, and the leaderboard for each line of work's own board.

What does not work?

• Keyword stuffing. The GEO paper measured it as worse than doing nothing, and search engines have long treated it as spam.
• Meeting a check instead of the reader. mttr once asked an answer to carry a number, and writers met it by turning "three ways" into "3 ways". The rule now checks only what code can honestly check, and leaves whether a paragraph answers to a person.
• Text written for a model and hidden from people, including instructions to assistants. Assistants that read the web treat it as untrusted, and search engines treat hidden text as spam.
• FAQ data for questions the page never shows. It is rare, and it is the kind of mismatch that gets markup ignored.
• Expecting a file to do the work. An llms.txt is cheap and harmless, and nothing major has said it reads one. See the llms.txt guide.

How do you check a page for GEO?

Read your site with mttr: it fetches the pages the way a crawler does and holds each to the same rules counted above, then lists the fixes worth the most. The free schema markup checker and heading checker look at one page, and the AI crawler checker reads your robots.txt.

Make a page easy to quote

• The words arrive in the HTML, before any script runs.
• At least one heading asks the question a reader would type.
• The paragraph right under it answers in 25 to 90 words, without leaning on the paragraph before.
• Figures that are compared sit in a table.
• The page shows when it was updated, and its structured data says so too.
• Sources a reader can follow are linked where the facts are.
• Nothing on the page is written for a model and hidden from people.

Sources

How the figures were counted. Every figure on this page is counted from the 316,710 sites mttr has read, each at its latest read, with drafts left out, as of October 26, 2026, and worked out again as sites are read. A share of pages counts only pages that were read: a site read only in part speaks for the pages read and no others. Each rule is the one in mttr's rulebook, decided by code. A score or a rule passed says pages are well formed. It does not say an assistant cites the site. The methodology explains each count.