A practitioner account of generative engine optimization: what the underlying research actually says, the ten changes that matter in priority order, and the before-and-after numbers from implementing all of them on this site.
Quick Answer
What is generative engine optimization (GEO)?
Generative engine optimization is the practice of structuring content so AI answer engines such as ChatGPT, Perplexity and Google AI Overviews quote it when answering a question. It differs from SEO in target: SEO competes for a position in a ranked list of links, GEO competes for inclusion in the synthesized answer itself. The peer-reviewed basis is Aggarwal et al., "GEO: Generative Engine Optimization" (KDD 2024), which found that adding citations, quotations and statistics raised a source's visibility in generated answers by up to 40%, while keyword stuffing reduced it.
What generative engine optimization actually is
Generative engine optimization (GEO) is the practice of structuring a page so that an AI answer engine — ChatGPT, Perplexity, Google AI Overviews, Claude — will quote it when answering a question. It is not a rebrand of SEO. Classic SEO competes for a position in a list of links; GEO competes to be the sentence the model repeats. A page can rank first on Google and never be cited once.
The distinction matters commercially because the citation, not the ranking, is what the buyer sees. When an AI assistant answers "which affiliate agency should I hire" it names two or three firms and the user rarely clicks through to check. Being absent from that answer is invisibility, regardless of your position on page one.
The evidence base is thinner than the marketing suggests
Most GEO advice traces back to a single peer-reviewed source: Aggarwal, Murahari, Rajpurohit and colleagues, "GEO: Generative Engine Optimization", published at KDD 2024 by researchers at Princeton, Georgia Tech and the Allen Institute for AI. It tested nine optimization methods against a benchmark of real queries and found that adding citations, quotations and statistics raised a source's visibility in generated answers by up to 40 percent, while keyword stuffing reduced it.
Two caveats are usually dropped when that 40 percent figure is repeated. It was measured on a custom generative engine built for the study, not on ChatGPT or Perplexity in production. And it is an upper bound across query categories, not an expected result. Treat it as directional evidence that structure and sourcing matter, not as a forecast.
What we changed on this site, and what it cost
Rather than describe the theory, here is the implementation we ran on xark.io in September 2026, with the numbers before and after. Every figure is reproducible from the site's own build output.
Meta descriptions. 108 of our resource pages were shipping the identical description "Xark — Affiliate Marketing Growth Agency." A prerender fallback was overwriting each page's own tags. Distinct descriptions went from 307 to 415 of 415 pages.
Titles. The longest page title was 185 characters, of which 42 were a repeated site-name suffix. Google truncates around 60. We cut the suffix to 7 characters and trimmed subtitles from over-long headlines, keeping the full headline as the on-page H1. Titles over 60 characters fell from 976 of 982 to 207.
Extraction blocks. We added a Key Takeaways block to 254 articles — four or five self-contained declarative sentences at the top of the page, mirrored into ItemList structured data. This is the single change most directly supported by the KDD paper, because it puts quotable, fact-dense sentences where a model reads first.
Navigation. Our own navbar contained three service links that returned 404 on every page of the site, the footer contained six more, and 22 of our 27 industry pages had no internal links pointing at them at all. A page an AI crawler cannot reach cannot be cited. Broken internal links went to zero.
Dead entity references. Our Organization schema listed a LinkedIn company page and an X handle that both returned 404. Pointing an engine at a non-existent profile weakens entity resolution rather than helping it, so we removed them.
The checklist, in the order that matters
- Make every page reachable. Audit internal links for 404s and redirect hops before writing anything new. Crawl budget spent on broken links is budget not spent on your content.
- Give every page a distinct title and description. Duplicates are a signal that the pages are interchangeable, which is a reason not to cite any of them.
- Lead with an extractable answer. Put a direct-answer block or key-takeaways list in the first screen. Write complete sentences that survive being lifted out of context.
- Cite real sources and name them. Link the primary source rather than gesturing at "research." This is the highest-leverage finding in the KDD study.
- Include specific numbers. Benchmarks, ranges and dated figures are quotable in a way that adjectives are not.
- Add FAQPage and Article schema. Structured data is not a ranking factor for AI engines, but it removes ambiguity about what the page asserts.
- Keep an llms.txt. It is a low-cost, machine-readable index of what your site is for. Ours lists every service, comparison and resource cluster.
- Allow the AI crawlers. GPTBot, ClaudeBot, PerplexityBot, Google-Extended and OAI-SearchBot should be explicitly allowed in robots.txt. Cloudflare began blocking several by default in 2024, so verify rather than assume.
- Fix your entity graph. Organization schema, sameAs profiles that actually resolve, and consistent naming across the site.
- Publish something only you can publish. Operator data, real benchmarks, honest comparisons. Generic explainers are already covered by Semrush and HubSpot, and you will not outrank them on their own ground.
What this does not do
None of the above manufactures authority. If nobody links to or mentions a brand, AI engines have little corroborating evidence that it exists, and structural work alone will not fix that. The realistic sequence is: make the site machine-legible first, because it is cheap and fully under your control, then build the external signals, which are slow and are not.
We publish our own numbers here partly to be useful and partly because a page arguing for source citation would be self-refuting without them.
Frequently asked questions
Is GEO different from SEO? They overlap in fundamentals — crawlability, clear structure, real expertise — and diverge in target. SEO optimizes for position in a ranked list. GEO optimizes for inclusion in a synthesized answer. A page can win one and lose the other.
Does llms.txt actually work? There is no public evidence that any major AI engine currently reads llms.txt as a ranking input. It costs almost nothing to publish and makes your site's structure explicit, which is why we maintain one, but treat claims of measured lift from llms.txt alone with scepticism.
How long until AI citations appear? Structural changes are re-crawled on the same cadence as normal indexing, so weeks rather than days. Citation behaviour also shifts when models are updated, which means measurement needs a baseline and repeat sampling, not a single check.
How do I measure AI visibility? Query the engines directly with the prompts your buyers would use, record whether you are named and in what position, and repeat weekly. There are commercial tools for this, but a spreadsheet and a fixed prompt list is a valid starting point.
