Service · GEO audit
In brief
A GEO audit — an audit for Generative Engine Optimization (GEO) — analyzes how visible your brand is in the answers of ChatGPT, Perplexity, Google AI Overviews, Gemini and Claude.
It measures, at prompt and engine level, how often your brand is mentioned, recommended and backed by a source. Repeated runs separate single hits from more stable patterns and show which competitors and sources appear instead.
In addition, the audit checks technical access, content extractability, brand entity and gaps in the source ecosystem. The result is a documented baseline with a prioritized roadmap and a plan for re-measurement.
Five engines, five different answers to the same purchase question. Checking only one engine shows only a fragment. A single hit proves nothing; the measurement becomes reliable only through repeated runs and points in time.
12–21 %overlap of the cited domain pools between the engines in a portfolio study of eight B2B projects.
Six areas — from demand and retrieval through technology and entity to citations, competition and business impact.
Not a slide deck that no one opens three months later. Six deliverables your team works with directly.
Calculate the economic potential of your AI visibility with transparent assumptions and three scenarios.
Calculate the business caseFive steps, typical timeframe 2–4 weeks from kickoff. For multiple markets, extensive competitor sets or large websites, the timeframe is extended on a binding basis in advance.
The foundation is your CRM: won and lost deals, sales conversations, support requests. I group the prompts into topic clusters, along the questions your customers use to begin a purchase decision.
All five engines, multiple runs, against a competitive benchmark.
Crawler access, structured data, entity consistency.
The findings move into the prioritization matrix, sorted by impact and effort.
Results presentation and action plan that your team or an agency can continue working on directly.
The most important step is the prompt set. The right prompt set comes from your CRM, not from gut feeling. How I build prompts is shown in the prompt library.
The starting point
GEO audit. Stands on its own. You get clarity about your gap and a plan you can implement yourself. Fixed package, fixed timeframe.
GEO program. Measure, prioritize and implement on a monthly cadence, with a competitive benchmark and C-level report. The outcome is an internal GEO knowledge base.
After the audit
Your team receives the backlog, templates and rules for re-checking and works through the roadmap independently.
You run a handover session and answer the technical questions of your current agency.
One-off support or a GEO program with monthly measurements.
Technical deep dive
Methodology, technical checks, metrics and limits — complete and traceable.
A single run is not reliable. The measurement becomes reliable only as an occurrence frequency over many runs, repeated over time.
The same question often produces different answers from a language model, governed by temperature and sampling. A single hit therefore proves nothing. Neither does a single miss.
Each prompt runs multiple times per engine over a defined period. What is counted is the frequency with which a brand appears, is recommended or cited — only a stable rate is a signal.
Every measurement carries its date, because the engines change. A value from today is no promise for next month; only ongoing monitoring makes the movement visible.
Illustrative simulation
"Name the top 3 most reliable bean-to-cup coffee machines for home use"
Run #1. Every run can flip. A single hit is not a reliable signal.
The number 100 and the values are for illustration only. The actual measurement scope is defined by prompt set, engine, market and the desired reliability.
A buyer types one question, but the AI engine rarely searches just once. It breaks the prompt into sub-queries and fetches its own sources for each: product, comparisons, prices, constraints, region and alternatives. This process is called Query Fan-Out.
Your page can fit the original question and still be missing from individual sub-queries. The audit therefore checks for which sub-queries you are retrieved — and for which you are not. How a prompt fans out is something I show in Query Fan-Out and sub-queries.
Six areas, from demand to prioritization. Each provides part of the answer to why your brand appears in AI answers or is missing.
Which questions your buyers really ask, from real sources instead of a keyword tool: real customer questions from CRM, sales and support, the purchase phases of the customer journey, market and language, branded and unbranded prompts.
Which sub-queries a prompt triggers and whether your page is retrieved in the process: likely sub-queries per prompt, comparisons and alternatives, geographic queries, verification of retrieval.
Whether an AI crawler reaches your page and can cleanly extract the answer: robots.txt, server rendering instead of pure JavaScript, semantic HTML and structured data, extractability of individual facts.
Which topics the models associate with your brand and which external sources carry that: associated categories, products, markets and competitors, external confirmation, gaps in the source ecosystem.
Where you are cited, mentioned or passed over: Citation Rate, Mention Rate, Recommendation Rate, AI Share of Voice, citations without a brand mention, mentions without a link. Measured with Peec.ai, the raw data stays traceable.
What comes first, measured by the impact of the problem, the effort to fix it, priority and possible effect on leads and revenue.
From the question to the recommendation, your brand passes through several stages. The audit finds the stage where the chain breaks.
| State | Meaning | To check |
|---|---|---|
| Not in retrieval | The engine did not find or did not choose your page | Access, relevance, indexing |
| Retrieved, but not cited | Page found, lost against other sources | Quality of the passage, evidence, authority |
| Cited, but brand not mentioned | Content used without attributing your brand | Entity Clarity, brand references |
| Mentioned, but not cited | The model knows the brand but does not use your page | own content and external confirmation |
| Mentioned and cited | full visibility | scale the successful pattern |
An AI agent does not see your design but the floor plan of your website. If one stage breaks, the chain ends before the citation.
| Checkpoint | Question | If it is missing |
|---|---|---|
| Crawler access | Do GPTBot, ClaudeBot, PerplexityBot and Bingbot get through robots.txt? Does a firewall block them? | The page cannot be retrieved via this access path |
| Server HTML | Is the content in the HTML or only after JavaScript rendering? | What the crawler does not render does not exist for it |
| Semantic HTML | Real headings, real tables, clean structure? | The model does not find the answer in the text |
| JSON-LD and Schema.org | Are your entity data machine-extractable? | Facts about your brand remain unreadable |
| Load time, Core Web Vitals | Does the page load fast enough for the crawl? | A precondition, not an end in itself |
A simplified reference model of a Retrieval-Augmented-Generation pipeline — a typical sequence, not a vendor-specific architecture.
A Ghost Entity becomes a Referenced Entity: a source the model knows and names.
Checked are the brand name, category, products, markets, competitors and leadership profiles as well as the consistency of the entity data — Named Entity Recognition and Entity Disambiguation included.
What matters is whether independent sources confirm the same facts and clearly connect the brand with its category. For this I measure the semantic proximity of your content to the purchase prompts.
Industry publications, comparison portals, independent reviews, specialist directories, communities, partner sites, marketplaces, LinkedIn, Wikidata and competitor sites shape the answers. For each type, the audit checks whether your brand appears and how it is described.
| Source | Use | Problem | Action |
|---|---|---|---|
| Comparison portal | basis of the recommendation | brand missing | outreach and inclusion |
| Industry magazine | confirms the expertise | outdated description | update the data |
| shapes the perception | frequent complaint without a response | correct the product fact | |
| confirms the entity | company name differs | harmonize the data |
The more complete the input data, the sharper the findings. If something is missing, we record it as its own uncertainty.
Missing data is documented as its own uncertainty.
The working definitions are applied consistently across all engines and runs and documented with the date of collection.
| Metric | Definition | Reference base |
|---|---|---|
| Mention Rate | Share of runs of a prompt set in which your brand is named in the answer — regardless of whether a source is linked. | runs per engine |
| Recommendation Rate | Share of runs in which your brand is not just mentioned but recommended as an option or included in a shortlist. | runs per engine |
| Citation Rate | Share of runs in which at least one of your URLs is cited or linked as a source of the answer. | runs per engine |
| Citation Share | Share of your cited sources among all sources cited in an answer — how large your share of the evidence drawn upon is. | cited sources per answer |
| AI Share of Voice | Share of your brand among all brands mentioned across a prompt set, averaged over the runs — your presence relative to the competition. | mentioned brands per prompt set |
The most important prompts, the search intents behind them and the sub-queries AI systems generate from them.
Where your page is retrieved, where it is cited and where your brand is mentioned, per engine. Citation Rate and Citation Share serve as the baseline for everything that follows.
Which entities your brand is linked to and which external sources feed the AI answers.
Which pages already deliver well-extractable answers, facts and evidence and which need to be reworked for that.
For each measure, the problem, the affected pages, the expected effect, the effort, the priority, the responsible role and the implementation mode. Sorted by the 3-level framework: Presence, Readiness, Business Impact. What the Business Case for AI Visibility roughly estimates in advance becomes a concrete sequence here.
We go through the findings together, explain the measurement limits and prioritize the next steps. Your team clarifies methodological, technical and organizational questions directly.
According to the G2 AI Search Insight Report 2026, 51 percent of the B2B software buyers surveyed say they begin their software research more often in an AI chatbot than in Google. The figure describes the behavior of the respondents and is not a conversion factor between Google search volume and the number of AI prompts.
The calculation is a scenario-based model, not a measurement of the actual prompt volume and not a revenue forecast.
A GEO audit rarely lands on just one desk. Three roles read the report differently.
CMO / Management
AI visibility per engine and purchase phase, expressed in Citation Rate and Citation Share, plus the line to the pipeline: which gap costs you inquiries — prepared so you can defend a budget for it at board level.
Marketing / SEO
For which question ChatGPT names three competitors and not you, which topics are missing, which third-party sources the models cite. The prioritization matrix becomes the work plan for briefings and tickets.
IT / Development
Blocked crawler access in robots.txt, content only after JavaScript, missing or faulty JSON-LD, load times above the threshold — every point concrete enough for the backlog.
I decline in advance if the audit brings you nothing. That saves us both time.
Without a solid base of content, the substance a model could refer to is missing. Then the first step is content, and the measurement comes afterwards.
Without sales knowledge, the prompt set cannot be aligned with real purchase decisions. An audit built on guessed prompts is not worth the price.
No one controls the output of a language model directly. I measure, diagnose and shift probabilities with evidence, and I say openly where the limit lies.
If the crawler cannot get through and the page only renders after seconds, that is first an SEO project on the foundation before GEO takes hold.
Both audits measure visibility, but in different spaces.
| Criterion | SEO audit | GEO audit |
|---|---|---|
| Main goal | Find the causes that prevent positions and organic traffic in classic Google search. | Find the causes why a brand is not mentioned, recommended or cited in the answers of generative systems. |
| What exactly is checked | The classic results list and the ability of the website to rank for target search terms. | The generated answers to real purchase questions and the presence of the brand in them. |
| Guiding question of the audit | Why does the desired page not rank? | Why does the brand or page not land in the answer and sources of the AI system? |
| Unit of analysis | Search term, position and target URL. | Prompt, single run, AI engine, brand mention and citation or source. |
| Core metrics | Positions of the target pages, impressions, clicks, CTR, organic traffic, indexability. | Mention Rate, Citation Rate, Citation Share and AI Share of Voice, separately per engine. |
| Measurement environments | Primarily Google Search, and further classic search engines if needed. | ChatGPT, Perplexity, Google AI Overviews, Gemini and Claude; Copilot can be added separately. |
| Data and tools | Google Search Console, crawling, server logs, Ahrefs or Sistrix, link and competitive analysis. | Repeated prompt measurements across several engines, Peec.ai and the raw answers, a competitive benchmark at prompt level. |
| Handling fluctuating results | Positions and clicks are viewed as a time series, yet the same search term usually returns a comparatively stable results list. | A single answer does not count as evidence. Each prompt runs multiple times per engine; what is measured is the frequency of appearance and the change over time. |
| Content focus | Fit to search intent, one search term per URL, eliminating cannibalization and duplicates, optimizing title and description, updating or merging weak pages. | Extractability of the answer: clear semantic structure, self-contained passages, facts, statistics, comparisons, currency and content proximity to real purchase prompts. |
| Role of the Google position | The position is one of the central outcomes of the optimization. | A high position does not guarantee a citation in a chat engine, and a lower-ranked page can be chosen as a source just as well. In an Ahrefs study of 15,000 prompts, the URL overlap of sources cited by ChatGPT, Gemini, Copilot and Perplexity with the Google top 10 averaged around 12% (Claude not studied). In the seoClarity sample, 94% of the AI Overviews examined contained at least one source from the organic top 20 — not evidence that 94% of all citations come from there. |
| Authority | Internal linking, external links, quality of referring domains and distribution of link equity are analyzed. | Brand mentions, the independent sources that support the brand, entity consistency, Knowledge Graph and Wikidata, and the citation graph are analyzed. |
| Mentions versus links | Backlinks and referring domains remain a separate object of examination. | Unlinked mentions count as their own signal of AI visibility. In the material presented, the correlation is 0.664 for brand mentions versus 0.326 for Domain Rating. This is the result of a single dataset, not proven causation and not a universally valid factor weight. |
| Technical foundation | Crawling, indexing, robots.txt, canonical and noindex, status codes and redirects, XML sitemap, crawl budget, rendering, mobile version and Core Web Vitals. | Access for AI and search bots, WAF and robots.txt, server-side HTML, semantic HTML, extraction and chunking of content, structured data, load time and technical reliability. |
| Special case of the Google AI features | An SEO audit is especially important when the problem concerns the Google AI Overviews: this feature is closely tied to the classic index and the results list. | There is no separate technical standard for the Google AI Overviews; Google recommends the same SEO fundamentals. The GEO layer here serves to analyze the actual appearance in answers and citations, not as a replacement for SEO. |
| Checking the brand entity | Can be part of E-E-A-T and structured data, but is rarely the central axis of the audit. | One of the core blocks: a consistent name, entity recognition, disambiguation, confirmation of the facts by external sources and the connection of the brand with its category. |
| Competitive analysis | Comparison of positions, content gaps, link profiles and site architecture. | Comparison at prompt level: whom the engine names instead of the brand, whom it cites and which third-party sources shape the answer. The results are assessed separately per engine, because their source pools differ significantly. |
| Typical result artifacts | Prioritization matrix, technical ticket backlog and KPI baseline for positions, impressions, clicks and index coverage. | Prioritization matrix by Presence, Readiness and Business Impact, roadmap and KPI baseline for Citation Rate and Citation Share per engine. |
| When to conduct it | Before a relaunch, after a traffic drop or major Google update, with stagnating positions or a loss of visibility against a competitor. | When buyers already use AI systems to make their choice, the website has a content base and sales or CRM provide real purchase prompts. |
| Limits | Meeting all requirements guarantees neither indexing nor a high position. | A "first place in ChatGPT" cannot be guaranteed: the output is stochastic, the engines change, and the audit measures and increases the probability of presence but does not control the answer directly. |
| When to start | When the website is poorly indexed, technically unstable, losing classic traffic or missing from the Google AI Overviews. | When the SEO foundation already holds but the brand is missing from ChatGPT, Perplexity, Claude or Gemini, or when the gap between ranking and citation is to be measured. |
Fixed package, fixed timeframe. The terms depend on the scope of the prompt set and the number of markets checked, and I share them with you on request. If you want to estimate the business lever in advance, the Business Case at eullrich.com/en/business-case runs three scenarios in two minutes.
An SEO audit checks the ranking factors of the classic results list. A GEO audit checks whether your brand is mentioned and cited in the generated AI answer. The URL overlap differs: in an Ahrefs study of 15,000 prompts, around 12 percent of the URLs cited by ChatGPT, Gemini, Copilot and Perplexity also appeared in the Google top 10 (Claude was not part of the study). In the seoClarity sample of 362,000 keywords, 94 percent of the AI Overviews examined contained at least one source that also ranked in the organic top 20; that does not mean 94 percent of all citations come from the top 20. A GEO audit measures exactly this difference.
The measured engines are ChatGPT, Perplexity, Google AI Overviews, Gemini and Claude. Copilot draws on the Bing index, and the technical check covers Bingbot access as well. On request, I add Copilot to the prompt panel.
No. For Google AI Overviews and AI Mode, the fundamental SEO requirements still apply; Google requires neither a special AI file nor any particular Schema.org markup (Google guide for generative search features). Structured data can describe entities and content more unambiguously for machines, but it must match the visible page content. Whether other AI systems fetch or consider an llms.txt has to be verified against their respective documentation and your own server logs. The GEO audit therefore first assesses content substance, technical access, retrieval, sources and brand entity — not individual supposed GEO shortcuts.
The typical timeframe is 2–4 weeks from kickoff. For multiple markets, extensive competitor sets or large websites, it is extended on a binding basis in advance.
For B2B brands and e-commerce in the DACH region whose buyers increasingly begin their decision in ChatGPT, Perplexity or the AI Overviews. The audit becomes worthwhile once there is an existing website with a solid content base, because otherwise there is little to measure.
Answer Engine Optimization (AEO) and Generative Engine Optimization (GEO) aim at the same outcome: appearing in the AI answer instead of only in the results list. GEO is the more common term and explicitly includes optimization for generative systems such as ChatGPT and Gemini. I work under the term GEO.
Write to me and I'll get back to you with the terms for your prompt set and your markets.