Generative engine optimization in 2026: what actually works
By Roshan, founder of SearchMaxx - - updated - 8 min read
Generative engine optimization (GEO) is the work of getting your pages cited when an AI engine answers a question. What works in 2026 is mostly what the original research found: pages that are easy to retrieve and full of specific, checkable facts get quoted, and keyword tricks do not. Most of the rest of what gets sold as GEO, from llms.txt files to "AI-friendly" rewrites, has no evidence behind it.
Our position is simple. GEO is not a new discipline that replaces SEO. It is SEO with a different finish line. The engine still has to find your page first. After that, your page has to be the one it chooses to quote.
Where the term comes from
The term comes from a paper called GEO: Generative Engine Optimization, first posted in November 2023 and published at KDD 2024, the ACM data mining conference held in Barcelona in August 2024. The authors are Pranjal Aggarwal (IIT Delhi), Vishvak Murahari, Karthik Narasimhan and Ameet Deshpande (Princeton), and Tanmay Rajpurohit and Ashwin Kalyan.
The setup was clean. The team built GEO-bench, a set of 10,000 queries drawn from nine datasets. For each query, a simulated engine took the top five Google results and had gpt-3.5-turbo write an answer that cited them. Then the researchers rewrote one source at a time using nine different methods and measured how much more of the answer came from that source.
They used two measures:
- Position-adjusted word count. How many words of the answer come from your source, weighted so that being quoted early counts more.
- Subjective impression. A model's judgment of how relevant, influential and prominent your source is in the answer.
The abstract's headline claim is that GEO "can boost visibility by up to 40%" in generative engine responses.
What the paper found
| Method | What it means | Result in the paper |
|---|---|---|
| Cite sources | Add references to credible sources | Among the top three methods |
| Quotation addition | Add relevant quotes from named people | Among the top three methods |
| Statistics addition | Replace vague claims with numbers | Among the top three methods |
| Fluency optimization | Make the text read more smoothly | Strong, and best when combined with statistics |
| Keyword stuffing | Repeat the target query's words | "Little to no improvement", and worse than doing nothing on the main measure |
In the authors' words, the top three methods "achieved a relative improvement of 30-40%" on position-adjusted word count and "15-30%" on subjective impression. The best pair, fluency plus statistics, beat any single method by more than 5.5%.
Three more findings matter for a software company:
- Lower-ranked pages gained the most. Citing sources "led to a substantial 115.1% increase in visibility for websites ranked fifth" in the search results. If you are not the top result, specific evidence on the page can still get you quoted.
- What works depends on the topic. Statistics helped most on law, government and opinion questions. Quotations helped most on people, society and history. There is no single trick for every page.
- A real engine moved less. On Perplexity, quotation addition was the best method, at 22% over the baseline on word count. Keyword stuffing did 10% worse than leaving the page alone.
What has held up
The paper is almost three years old, and it tested a simulated engine on an older model. The engines have changed a lot since. Some of its conclusions have held up well, mostly because they match what the engines now say about themselves.
You have to be retrieved before you can be cited. In the paper, every candidate page came from the top five Google results. That design choice turned out to describe the real world. Google says a page must be "indexed and eligible to be shown in Google Search with a snippet" to appear in AI Overviews or AI Mode (Google Search Central, AI features and your website, updated December 2025). OpenAI says sites that block its search crawler "will not be shown in ChatGPT search answers" (OpenAI crawler docs). Ahrefs found that 76% of pages cited in AI Overviews ranked in Google's top 10 for the query (Ahrefs, July 2025). We cover the full ranking question in Is SEO dead in 2026?
Specific facts beat vague claims. Take Paidwick, an example invoicing tool. A sentence like "Paidwick cuts invoice processing time" gives an engine nothing to quote. A sentence that names a number, how it was measured, and on what data does. The paper's three best methods all add something checkable: a number, a named source, or a quote. That is also what makes a page useful to a buyer, which is why we think it has held up.
Keyword tricks do not work. Keyword stuffing was the only method that made things worse in the paper. Google now names the related tactic directly: making pages mainly to manipulate AI responses "violates Google's scaled content abuse spam policy" (Google Search Central, optimizing for generative AI features, updated July 2026).
What is folklore
A lot of GEO advice in 2026 has no study behind it, and some of it contradicts what the engines publish. Here is what we tell clients to skip or treat with caution.
llms.txt as a ranking lever. Google is direct about this one. Its optimization guide says "You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search," and that Google Search does not use llms.txt. An llms.txt file is cheap and harmless. We have one. It is not why anyone gets cited.
Special schema for AI. Google says there is "no special schema.org markup you need to add" for AI features. Structured data still helps with rich results and keeps facts consistent, so keep it. Do not expect a new schema type to get you into AI answers.
Chopping pages into chunks. A popular claim says engines prefer pages broken into tiny self-contained blocks. Google's guide says "There's no requirement to break your content into tiny pieces for AI to better understand it." Clear headings help readers and engines alike. Turning a guide into fifty disconnected fragments helps nobody.
Writing in a special style for AI. Google: "You don't need to write in a specific way just for generative AI search." The paper agrees in its own way. Its best methods are just good writing habits: cite, quantify, quote.
Buying mentions. Some advice says to seed your brand name across forums and listicles so models pick it up. Google says seeking inauthentic mentions "isn't as helpful as it might seem." It is also the fastest way to lose the trust of the communities that actually shape buyer opinion.
Tools that claim to see inside the engine. Google: "No third-party tool has access to our internal ranking or AI systems." Tools that track whether you appear in AI answers are useful. Tools that claim to know why are guessing.
"GEO replaces SEO." Google's guide says optimizing for generative AI search "is optimizing for the search experience," and so it is still SEO. Other engines are less explicit, but the retrieval evidence above points the same way.
One caution about our own sources. Google's guidance covers Google. ChatGPT, Claude and Perplexity publish less about how they choose citations. Where we have no published data for an engine, we say so rather than guess.
A practical priority list for 2026
This is the order we work in. Each step depends on the one before it.
| Priority | What to do | How to check it |
|---|---|---|
| 1 | Let the AI search crawlers in | robots.txt allows OAI-SearchBot, Claude-SearchBot, PerplexityBot and Googlebot. Your firewall does not block them. See which AI crawlers to allow. |
| 2 | Send real text without JavaScript | A plain fetch of each key page returns your product description and pricing. See AI crawlers don't run JavaScript. |
| 3 | Get indexed and rank for buyer questions | Your key pages are indexed in Google and Bing and rank on page one for the questions buyers ask while choosing. |
| 4 | Answer one question per page, directly | The first two or three sentences answer the question in the title. A reader could quote them alone. |
| 5 | Add specific evidence | Replace vague claims with numbers, named sources and short quotes. Show where each number came from. |
| 6 | Keep your facts consistent | Pricing, features and integrations say the same thing on your site, your docs and your profiles elsewhere. |
| 7 | Earn real mentions | Get named in places buyers already read: comparisons, reviews, forums, and docs of tools you integrate with. |
| 8 | Measure | Track Google clicks and impressions in Search Console, and ask the engines your buyers' questions on a fixed schedule. |
Steps 1 to 3 are where most sites fail. In our experience, the common reason a software company is missing from AI answers is not its prose. It is a robots.txt rule, a JavaScript-only page, or a page that does not rank anywhere.
Steps 4 and 5 are where the GEO paper's findings apply directly. In the case study we published, most of the growth came from pages built for questions buyers ask while choosing, such as "best model for coding" guides. Pages like that state a verdict and back it up, which is the kind of page AI engines cite.
What we still do not know
We want to be clear about the gaps.
- No public study has repeated the GEO experiment on current engines at scale. The original tested
gpt-3.5-turboand a small Perplexity sample. Newer engines search more widely. Google says AI Overviews and AI Mode may run several related searches at once, which it calls "query fan-out", so the set of candidate pages is larger than five. - Visibility in an answer is not the same as a visit. The paper measured how much of the answer came from a source. It did not measure clicks or sales. Studies on clicks tell a more mixed story, which we cover in Google AI Overviews and AI Mode in 2026.
- Each engine is different. A method that helps in one engine may do nothing in another. The paper found the effect varies by topic, too.
So test on your own pages. Pick ten buyer questions. Record which engines name you today. Make one change at a time and check again in a few weeks.
Where to start
If you do one thing this week, check steps 1 and 2. They take minutes and they decide whether any of the rest matters. Our free check tests whether the AI search crawlers can reach and read your site in about 10 seconds.
Want this done for your site?
Every day we make one change to your website that helps AI engines name you, and a person checks each one. Email me your website and the question your buyers ask AI.