Need Help? We’ve Got You Covered.
About GEO itself
Generative Engine Optimization is the practice of structuring a website so AI assistants can reach it, extract passages from it, and cite it in their answers. It covers crawler access, machine-readable structure, and whether the wider web describes your brand consistently enough for a model to resolve it. The term comes from a 2024 Princeton and Georgia Tech published at KDD 2024, which tested nine optimization methods across roughly 10,000 queries.
It works, and there is peer-reviewed measurement behind it. The Princeton and Georgia Tech study, published at KDD 2024, tested nine optimization methods across roughly 10,000 queries and measured visibility improvements of 22% to 41%. Citing sources, adding statistics and adding quotations were the strongest levers. Keyword stuffing was the weakest and measurably reduced visibility. Lower visibility sites gained the most.
Because of what it converts at. Semrush measured the average AI search visitor as 4.4 times as valuable as the average organic one across 500+ topics. Ahrefs reported that on its own site, half a percent of traffic produced 12.1% of signups, a 23x rate. In one Seer Interactive client case study, ChatGPT traffic converted at 15.9% against 1.76% for Google organic. Add the growth curve, Adobe recorded +1,200% to US retail sites between July 2024 and February 2025, and a small channel becomes the highest-value one most teams do not measure.
Largely yes. GEO, AEO, LLMO and AI SEO describe overlapping work under different names, and the distinctions rarely change what gets done to a site. If a provider sells them as three or four separate products, ask which specific tasks differ between them. We use one term and one scope to avoid selling the same work three times.
Between roughly a quarter and a half of queries, and every published count is rising. Conductor measured 25.11% across 21.9 million queries. BrightEdge measured 48% across nine commercial verticals in February 2026, up 58% in a year. The spread is the keyword panel each study used, so treat any single figure with suspicion, including ours. It runs highest on questions: Pew found AI summaries on 60% of queries starting with who, what, when or why, and on 53% of searches ten words or longer, against 8% of one-to-two-word searches.
Being cited means an assistant consulted your page. Being recommended means it put your brand on the buyer's shortlist. Citation is mostly won on your own site through structure and clarity. Recommendation is governed by what the rest of the web says about you: reviews, forums, press, analysts. Lily Ray analysed 100 B2B category queries in 2026 and found that in 224 of the 323 citations self-promotional guides earned, 69%, Google cited the brand's own page and recommended somebody else. Full explanation on the visibility ladder.
You are probably climbing the citation and mention rungs while the recommendation rung stays flat. Most tools report the first three and label the result "AI visibility". Recommendation is governed by web-wide consensus rather than by your own content, so rising citations alongside flat recommendations is a specific, diagnosable gap: your content works and the web does not yet corroborate it.
Probably not what is happening. Similarweb tracked real journeys for seven days after an answer and found 55.9% of AI-influenced traffic arrived through search, not as an AI referral, against 40.4% for visits with no AI influence. The assistant makes the recommendation, the buyer types your name into Google, and the visit lands as ordinary branded organic. If you judge this channel by the referral line, you see the smaller half of the effect. Watch branded search volume and add a "how did you hear about us?" field.
Crawlers and access
It depends on how you make money, and anyone answering instantly is not thinking. Cloudflare Radar measured crawl-to-referral ratios on 31 May 2026: ClaudeBot 10,300 to 1, GPTBot 903.8 to 1, Googlebot 5.2 to 1. If you sell advertising against pageviews, blocking is defensible. If you want to be recommended when someone asks an assistant, blocking removes you from consideration.
Allow the search-and-cite crawlers if you want citation: GPTBot and ChatGPT-User (OpenAI), ClaudeBot (Anthropic), PerplexityBot (Perplexity), Google-Extended (Gemini and AI Overviews), Bingbot (Copilot). You can separately block training-only crawlers such as CCBot. Every crawler identifies itself with its own user-agent, so this is a configuration decision rather than an all-or-nothing switch.
Yes, and for most commercial sites that is the right configuration. Named AI crawlers announce themselves and obey robots.txt. Price-scraping traffic mostly arrives disguised as a browser and ignores robots.txt entirely, so a blanket block stops the crawlers that follow the rules and does nothing to the ones that do not. Assistants also do not need live pricing to recommend a brand.
Not necessarily, and this is the most common false assumption we correct. Your CDN or WAF can refuse the same crawler your robots.txt invites. In our 2026 tour operator study, a market leader with the highest technical score in its field scored zero in practice because Akamai answered GPTBot with a 403. Testing needs a real fetch as each named agent.
Often yes. If your content is assembled in the browser, the HTML an AI crawler receives may contain no text at all. We measured one operator whose page returned 115,801 bytes, 75% of it JavaScript, with no title tag and no body text. An AI reader asked to summarise it returned the site's Google Analytics script. The fix is server-side rendering of key content, not a rebuild.
Doing the work
Four phases: measure, fix, build the entity, re-measure. That covers 78 itemised deliverables, from robots.txt crawler rules and JSON-LD across ten schema types to Wikidata entries, multilingual content and a monthly re-run against a frozen prompt set. The full list is on How It Works, numbered so you can count them rather than take our word for it.
No, and right now it is unusually early. Across sixteen major European operators we measured in 2026, the best score in the whole field was 68 out of 100 and fourteen had no llms.txt at all. Most sectors look similar. The position is available in a way it will not be once assistants have settled on who they cite, and every month of reinforcement makes displacement more expensive.
Because AI answers are probabilistic and a single reading is not a measurement. Schulte, Bleeker and Kaufmann measured source overlap in AI answers between consecutive days at 34% to 42%: roughly two thirds of the sources cited for a query one day differ the next. A snapshot showing you absent may just be the run where you were absent. Fixing the prompt set and the cadence is what turns readings into a trend.
Technical fixes get picked up in weeks. Entity and authority work compounds over months. Your first re-measurement lands 30 days after the baseline, which is the first point at which movement is provable. Anyone quoting a guaranteed number of days to a first citation is guessing, because nobody controls model behaviour or refresh cycles.
The deployment, yes, and we write the implementation guide for exactly that, naming the file, the location and the verification step. What internal teams are rarely set up for is the standing measurement: a fixed prompt set through several engines every month, recorded against history, showing which competitor took your citations. That is an ongoing job with a database behind it.
No. We can hand every artifact to your developers with an implementation guide, or deploy them ourselves if you grant access. The choice does not change what you receive. Either way the files and the measurement baseline are yours to keep.
Yes, and it is usually the right structure. Technical foundations overlap, and the extraction, entity and measurement layers are additive. We have no interest in displacing an agency that is doing good work on rankings, because rankings are still where most of the traffic is.
We guarantee the measurement, the implementation and a baseline that proves whether the work moved. What we will not do is guarantee an engine's output, because nobody controls that and any provider claiming otherwise is telling you something they cannot deliver. What we can tell you is that the levers are measured: the Princeton study found visibility gains of 22% to 41%, with the biggest gains going to sites starting from low visibility.
The questions vendors dodge
Yes, llms.txt and llms-full.txt ship as standard in every engagement, generated from your sitemap and kept current. Adoption sits at 10.13% of 300,000 domains, so having one puts you in a minority of sites that have thought about machine readability at all. Writing it is also a useful exercise in its own right: several clients found orphaned pages and duplicate sections while we built theirs.
No. Google's AI features guide states plainly that no special markup or files are required for AI Overviews or AI Mode, because those features run on core Search ranking. Structured data still helps materially with ChatGPT, Claude and Perplexity, and it helps your conventional SEO, which is why we ship it. It is not a Google AI requirement.
No. Google names AI-targeted content variants and chunking pages into fragments for AI as risks under its scaled content abuse policy. Write for the reader, then organise so a passage lifts cleanly: a direct answer under each heading, tables for comparisons, sources in the sentence. Those are normal editorial standards, not machine-targeted tricks.
Check the source and the definition, because they measure different things. Pew Research, tracking 68,879 real Google searches by 900 US adults in March 2025, found 8% click-through when an AI summary appeared against 15% when it did not. Similarweb puts overall zero-click Google searches at 68% in 2026. Those two numbers answer different questions and neither is "the" zero-click rate. Sites quoting "over 50%" and "over 60%" on different pages of the same site are quoting neither.
Because a measurement nobody can check is a claim, not a measurement. Our tour operator index names sixteen companies with a published methodology, so anyone can dispute a number on the merits. Every score comes from a public homepage. We never share one client's report with another, and any company listed can request their full detail.
Commercial
Per engagement, quoted after the free report. There is no rate card because there is no standard job: a five-page site in one language and a 400-page site across four markets are different amounts of work. The quote reflects the hours a named consultant and the team spend on it. Scope and price are agreed in writing before anything starts.
Because every engagement has a person assigned to it, and the hours change with the work. A published rate card would mean either overcharging small accounts or underserving large ones, and we would rather quote honestly than sort you into a tier. What determines your number is published: page count, markets, languages, engines tracked, current technical state, and whether we implement or hand over.
Yes, and in some sectors it is what we recommend. Tour operators and other high-volume businesses often prefer their costs to move with their season rather than sit as a fixed line, so we structure those engagements as a percentage. We also build performance-linked, milestone-based and retainer structures. Ask for the shape that fits how your business earns, and we will quote it that way.
Usually not, and the comparison is the wrong one. Platforms run roughly €29 to €500 a month and give you a dashboard: prompts tracked, a score, a trend line. Nobody fixes anything. Our engagement includes the measurement plus the implementation, the entity work and a named person doing it. If a dashboard is what you need, buy the dashboard, and we will tell you so on the call.
Nothing, and it carries no obligation. You get the technical readiness score, the effective score after crawler testing, the per-crawler results and the prioritised findings whether or not you engage us. Several companies have taken the report, handed it to their own developers and fixed the problems themselves. That is a legitimate outcome.
Any size where being on an AI shortlist affects revenue. The method is fixed and the scope is built per industry, market, audience, company size and country. A regional operator and a multinational group get the same measurement discipline and different prompt sets, languages and priorities.
Everything: the frozen prompt set, the history database, every generated artifact and the implementation guide. The baseline is yours to re-run with or without us. Holding a client's measurement hostage is a way to keep accounts that deserve to leave.
Swim in the spotlight.
Thysanos is the Generative Engine Optimization service of SyeniteLabs.
© 2026 · THESSALONIKI