What a GEO SEO agency actually does
A GEO SEO agency works on generative engine optimization: making your brand, products and pages more likely to be retrieved, cited and accurately described when people ask AI systems questions. Those systems include ChatGPT search, Perplexity, Google AI Overviews and AI Mode, Gemini and Microsoft Copilot. Classic SEO asks "where does our page rank?" GEO asks two more questions. When an AI answer is written about our category, are we in it? And is what it says about us correct?
The work is not magic, and it is not separate from search. Most generative engines ground their answers in documents they retrieve from the web or from a search index at the moment the question is asked. A page that can't be crawled, indexed or understood can't be retrieved. A brand that nobody else on the web mentions gives a model little reason to recommend it. So a good GEO programme is mostly disciplined SEO, content and digital PR, aimed at a different output and measured in a different way.
Two boundaries are worth stating up front. First, this page covers visibility inside generative answers. If you want the wider programme (technical SEO, content and links, built for search engines that now include AI features), see our AI SEO agency services. Second, "geo" in SEO sometimes means geographic targeting: ranking in particular cities or countries. That is a different discipline, covered by our local SEO and enterprise and international SEO work.
How generative engines pick the sources they cite
Most AI search products use some form of retrieval-augmented generation (RAG). The system turns a question into one or more searches, retrieves candidate documents, and has a large language model write an answer grounded in those documents. It often shows links to some of them as citations. Google describes a version of this for its own products: AI Overviews and AI Mode may use a "query fan-out" technique, issuing multiple related searches across subtopics to build a response, according to Google Search Central's guidance on AI features.
This has three practical consequences. You have to be reachable by the crawler or index each engine uses. You have to be retrievable for the sub-questions the engine generates, not only for the head term. And your content has to be usable: clear, specific and easy to quote, so that it supports the sentence the model is writing.
Access differs by engine, and the bot names are often confused. The table summarises what each provider documents publicly.
| Engine | What it relies on (as documented) | Key control | First-party reporting |
|---|---|---|---|
| Google AI Overviews and AI Mode | Google's Search index. A page must be indexed and eligible to show in Search with a snippet (Google). | Normal Googlebot access; snippet controls such as nosnippet and max-snippet. Blocking Google-Extended does not affect inclusion or ranking in Google Search (Google crawler docs). | Included in Search Console's Performance report under the Web search type, not broken out separately. |
| ChatGPT search | OAI-SearchBot surfaces websites in ChatGPT's search features. GPTBot is a separate crawler used for model training (OpenAI bot docs). | Allow OAI-SearchBot in robots.txt if you want to appear. Blocking GPTBot is a training decision, not a search one. | No publisher report documented by OpenAI (checked October 2026). Measure by sampling prompts and by referral traffic in analytics. |
| Perplexity | PerplexityBot surfaces and links websites in Perplexity results and, per Perplexity, is not used to train foundation models. Perplexity-User fetches pages for user-initiated requests (Perplexity bot docs). | Allow PerplexityBot in robots.txt and in any WAF or bot-management rules. | None. Measure by prompt sampling and referrals. |
| Microsoft Copilot and Bing AI answers | Content retrieved through Bing. Bing reports "grounding queries" used to retrieve cited pages. | Normal Bingbot access and Bing indexing. | Bing Webmaster Tools' AI Performance report (public preview since February 2026) shows citations, cited pages and grounding queries for Copilot, Bing AI summaries and some partners (Bing Webmaster Blog). |
A common risk is a security or CDN rule that quietly blocks AI search crawlers while the marketing team assumes the site is open. Checking server logs and firewall rules against these documented user agents is one of the first things we do.
What the GEO research found, and what it didn't
The term "generative engine optimization" comes from a paper by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, researchers from Princeton University, Georgia Tech, the Allen Institute for AI and IIT Delhi. It was first posted in November 2023 and published at KDD 2024 (arXiv 2311.09735). Many agency pages quote it loosely, so here is what it actually did.
- Benchmark: the authors built GEO-bench, a set of 10,000 queries drawn from sources including MS MARCO, Natural Questions and real Perplexity.ai queries.
- What was measured: how visible a source was in the generated answer. They used two impression metrics: Position-Adjusted Word Count (how much of the answer is attributed to a source, weighted by where it appears) and Subjective Impression (an LLM-judged score covering factors such as relevance and influence).
- What was tested: nine content edits, including adding statistics, adding quotations, citing sources, improving fluency, using an authoritative tone and keyword stuffing.
- What worked: Quotation Addition and Statistics Addition produced relative visibility improvements of roughly 30 to 40% on Position-Adjusted Word Count, and Cite Sources also helped. The headline "up to 40%" figure comes from this.
- What didn't: keyword stuffing offered little to no improvement. Effects also varied by domain, and lower-ranked sources tended to gain more than sources that already ranked first (in the paper, Cite Sources lifted fifth-ranked sources sharply while first-ranked sources lost visibility).
The caveats matter as much as the findings. These were controlled experiments on a research benchmark and a 2023-era engine setup, with gains measured relative to a baseline answer. They were not traffic or revenue outcomes. Commercial engines change often and don't publish their retrieval rules. We use the paper as a guide to which levers are worth testing: evidence density, attributable statistics, quotable expert statements and clear sourcing. We don't treat it as a promise that any page will gain 40%.
Our generative engine optimization services
Mediardx is an SEO agency based in India that works remotely with clients in the US, UK, Canada and Australia. Our GEO work is organised into six workstreams. Each one produces a concrete deliverable you keep, whether or not you continue working with us.
1. AI visibility baseline
We build a prompt set for your category (described in the measurement section below), run it across the engines that matter to your buyers, and record where you are mentioned, where you are cited, which competitors appear, and which third-party sources the engines lean on. This baseline sets the starting point that every later report compares against.
2. Crawler, index and rendering access
We check robots.txt, CDN and WAF rules against the documented AI search crawlers. We confirm that key pages are indexed in Google and Bing, are eligible for snippets, and render their main content without depending on client-side scripts. This overlaps heavily with a technical SEO audit, and it should, because retrieval starts with access.
3. Citable content engineering
Generative answers are assembled from passages. We rework priority pages so that each important question gets a direct, self-contained answer near the top of its section. Claims carry dates and sources, and original data or expert statements are easy to lift and attribute. Where the prompt research shows the engines fanning out to sub-questions you don't answer anywhere, we plan new pages with our SEO content team.
4. Entity and brand consistency
Models describe brands using what the web says about them. We audit how your organisation, products and people are described across your site, structured data, business profiles, directories, review platforms and partner pages. Then we fix contradictions: old names, wrong categories, outdated pricing or locations. The goal is accurate descriptions, not just more of them.
5. Structured data that matches the page
We implement schema.org markup (Organization, Product or Service, FAQ, Article and others where eligible) and keep it consistent with the visible content. Google's guidance is to keep structured data accurate to what is on the page. It also says that no special markup or "AI text files" are required to appear in its AI features (Google Search Central). So we treat schema as a clarity aid, not a GEO switch.
6. Third-party mentions and citations
Engines frequently cite third-party pages: reviews, comparisons, industry publications, forums, documentation. The baseline shows which of those sources appear for your prompts. We then plan whitehat ways to earn a place in them: digital PR built on original data, expert commentary, and accurate listings on relevant review and comparison sites. We don't buy placements or post fake reviews.
How we measure GEO: prompt sets, citation share and testing cadence
Measurement is where most GEO offers are weakest. AI answers are not a fixed ranking. In research run by SparkToro and Gumshoe and reported in January 2026, repeated runs of the same brand-recommendation prompts in ChatGPT and Google's AI produced the same list of brands less than 1% of the time (SparkToro). A single screenshot, or an "AI ranking position", tells you very little. What can be measured reliably is how often you appear across many prompts and runs.
Our method has four parts. The numbers below are the typical design choices we start from and adjust per client. They are not results.
- Build a prompt matrix. We cross your core topics with buying stages (problem, solution, comparison, vendor shortlist) and with realistic modifiers such as industry, company size, location and budget. A typical starting set is 50 to 150 prompts, written the way buyers actually phrase questions, not as keywords.
- Sample, don't snapshot. Each prompt runs several times per engine in each measurement cycle, with logged-out or clean sessions where the engine allows it, and the date, engine and mode are recorded. Repetition is what turns noisy answers into a usable rate.
- Score every answer. We record whether the brand is mentioned, whether your domain is cited, which competitors appear, which other domains are cited, and whether the description of your brand is accurate.
- Review on a fixed cadence. We take a baseline before any changes, track monthly, and review the prompt set quarterly so it keeps pace with how your buyers search. Changes we ship are logged against dates, so movements can be read in context rather than credited to one tactic.
| Metric | How it is calculated | What it tells you |
|---|---|---|
| Mention rate | Answers that name your brand ÷ total answers sampled | Whether you are in the engine's consideration set |
| Citation rate | Answers that link to your domain ÷ total answers sampled | Whether your own pages are used as sources |
| Citation share | Citations to your domain ÷ all citations across the sampled answers | Your share of the sources the engines rely on |
| Competitive share of voice | Your mentions ÷ mentions of you plus a fixed competitor set | Relative visibility against named rivals |
| Accuracy rate | Mentions with a correct description ÷ all mentions | Whether AI answers describe you correctly |
| Source mix | Most-cited third-party domains for your prompt set | Where to focus PR, listings and partnerships |
Alongside these sampled metrics we report first-party data where it exists: Google Search Console (AI feature traffic is included in the Web search type), Bing Webmaster Tools' AI Performance report, and referral sessions from AI assistants in your analytics. When a number isn't available, the report says so rather than estimating it.
GEO vs SEO: what changes and what stays the same
GEO doesn't replace SEO. It builds on it. Google says there are no additional requirements to appear in AI Overviews or AI Mode beyond the same SEO best practices that apply to Search generally (Google Search Central). What changes is the unit of competition and the way success is measured.
| Classic SEO | GEO | |
|---|---|---|
| Output you compete for | A ranked link on a results page | A mention or citation inside a generated answer |
| Unit of content | The page | The passage, fact or statement that can be quoted |
| Query scope | The keyword and its close variants | The question plus the sub-questions an engine fans out to |
| Main measurement | Rankings, impressions, clicks | Mention rate, citation share, accuracy across sampled prompts |
| Stays the same | Crawlability, indexation, helpful and specific content, a clear site structure, reputation earned on other sites | |
llms.txt, schema and other GEO myths
"You need an llms.txt file." llms.txt is a proposal published by Jeremy Howard in September 2024 for a Markdown file that summarises a site for language models (llmstxt.org). It is not an adopted web standard. Google's own documentation says you don't need to create new machine-readable files or AI text files to appear in its AI features (Google Search Central). It can be a reasonable low-cost addition for documentation-heavy sites. We don't sell it as a visibility lever.
"Special schema gets you into AI Overviews." No special markup is required. Accurate structured data helps machines understand a page, and that is reason enough to maintain it.
"Blocking AI crawlers protects you without side effects." It depends on which crawler. Blocking a training crawler such as GPTBot, or opting out with the Google-Extended robots.txt token (which covers Gemini training and grounding, not Google Search), is a separate decision from blocking a search crawler such as OAI-SearchBot or PerplexityBot. Blocking the search crawlers can remove you from those engines' answers.
"An agency can guarantee AI citations." It can't. Google states plainly that no one can guarantee a #1 ranking on Google, and generative answers are even less deterministic than rankings.
How to evaluate GEO SEO agencies
Lists of "top GEO SEO agencies" are mostly written by agencies. Whoever you shortlist, including us, these questions separate a method from a pitch:
- How do you measure visibility? Expect a defined prompt set, repeated runs and metric definitions. Be wary of a single "AI rank" number.
- Which engines, and which crawlers? The agency should know the difference between training crawlers and search crawlers, and check your firewall as well as robots.txt.
- What will you change on our site, and what off it? GEO without content and third-party reputation work is usually just reporting.
- How do you handle accuracy? Being described wrongly can be worse than not being mentioned.
- What do you own at the end? The prompt set, raw answer logs and dashboards should belong to you.
Red flags: guaranteed citations or rankings, llms.txt sold as the core deliverable, case-study percentages with no baseline or method, and any tactic that relies on fake reviews, spammy Q&A posting or hidden text aimed at models.
Timelines and what GEO can't do
A typical engagement starts with two to four weeks of baseline and access work, then monthly cycles of content, entity and PR work with measurement. Fixes to crawler access can show up quickly in engines that retrieve live. Changes that depend on third-party mentions and reputation take longer, often several months, and some categories are dominated by sources you can't easily displace. These ranges are typical, not guaranteed.
GEO also can't make an engine recommend a product that doesn't fit the question, and it can't fix a weak offer or poor reviews. Where those are the real constraints, the baseline will usually show it, and we'll tell you. For scope and pricing, contact us with your category and target engines. To see how engagements run week to week, read our process.