Generative engine optimisation (visibility in AI search assistants)

Customers increasingly ask an AI assistant "which provider should I choose?" instead of scrolling through ten blue links. Generative engine optimisation (GEO, also called AEO or LLMO) is the practice of measuring whether and how a brand appears in those answers, and of improving the sources the assistants draw on. AI itself is used on both sides: the engines are LLMs, and the monitoring tools use LLMs to run prompts at scale, classify mentions and suggest fixes.

Problem & value

• Today: marketing teams track rankings and clicks. In AI answers there is often no ranked list and no click, and answers vary between runs, models, languages and users. Brands do not know if they are recommended, described correctly or replaced by a competitor.
• AI value: automated, repeated sampling of many prompts across engines; classification of mentions, sentiment and factual errors; analysis of which third-party pages are cited; drafting of content and structured-data fixes for human review.
• Evidence base: the original GEO paper (Aggarwal et al., KDD 2024) reports that adding citations, quotations and statistics to content raised visibility in generative engine responses by up to about 40 % on its benchmark, while keyword stuffing helped little. These are research-benchmark results, not guarantees for production engines [1].
• Realistic framing: GEO overlaps strongly with good SEO. Google states that its AI features (AI Overviews, AI Mode) use the same foundational SEO practices and that no special markup is required to appear [2].

Variants

• Local SMB / hospitality: accurate entries in Google Business Profile, Bing Places, review platforms and maps data, which assistants reuse for "best X near me" questions.
• E-commerce: clean product feeds, schema.org Product data and reviews; some assistants offer product feeds or checkout integrations (see REC-001, KNOW-005).
• B2B / SaaS: comparison pages, documentation, analyst and community mentions that assistants cite for "best tool for ..." questions.
• Publishers and media: visibility versus licensing and crawling control (allow search crawlers, decide on training crawlers).
• Reputation and accuracy monitoring: detecting wrong claims about prices, opening hours, safety or legal facts, including for public bodies and non-profits.
• Private use: individuals checking what assistants say about them are a data-subject use, see GDPR notes below.

How it works

1. topic & prompt design (customer questions × market × language × persona)
2. scheduled sampling of engines (API or UI-based, N repeats per prompt)
3. answer parsing (brand/competitor mentions, citations, sentiment, claims)
4. source analysis (which domains and pages are cited; owned vs third party)
5. gap analysis (missing facts, weak pages, blocked crawlers, missing structured data)
6. action plan (content, PR/earned mentions, technical fixes)
7. human review
8. publish
9. re-measure (trend, share of voice, AI-referral traffic)

• Prompt set: 50-300 realistic questions built from search queries, sales calls and support tickets; an LLM can expand and cluster them, humans validate. Avoid prompts that name the brand only, since they overstate visibility.
• Sampling: answers are non-deterministic and personalised, so each prompt is run repeatedly and results reported as rates with ranges, not single ranks.
• Classification: an LLM-as-judge or small classifier labels mention, position, sentiment and factual correctness against a ground-truth fact sheet; spot-check by humans.
• Human-in-the-loop: people approve all published content, correct facts and PR outreach. Content is never auto-published at scale.

Data preparation

The data here are mostly public or marketing-owned: a prompt set of customer questions per topic and language, sampled answers and citations from AI assistants, your website content, structured data, product feeds and robots.txt rules. In an SMB the content sits in the CMS or shop system, Google Business profile and review portals; product facts may be scattered between the ERP, PDFs and brochures. Typical gaps are outdated or contradictory facts across channels, missing structured data, no list of real customer questions, and no baseline measurement.

Preparation steps: - Collect real customer questions from support mail, sales calls and search queries to build the prompt set (PREP-016). - Make one authoritative fact sheet per product, service and location, and fix contradictions (PREP-013). - Align terminology and synonyms (PREP-015). - Check structured data and make sure crawlers are not blocked unintentionally. - Record a baseline: run each prompt repeatedly, since answers vary.

Effort is low, roughly 15 to 25 percent of the work, mostly content clean-up. No personal data is needed, so GDPR effort is small.

Implementation paths

Which path for whom

• SMB (A): Dedicated GEO monitoring tools are affordable and standard; no own data integration is needed.
• Enterprise (A): This is a commodity monitoring function, and tools cover multi-market prompt tracking; add B only for custom reporting.
• Developer (B): Building your own prompt sampling and citation tracking via model APIs gives control over prompt sets and data.

Path A — Buy (recommended for most SMBs)

• AI-visibility trackers: Profound, Semrush AI Visibility / AI toolkit, Ahrefs Brand Radar, Similarweb AI tracking, Peec AI (Berlin, EU), OtterlyAI (Austria, EU), Scrunch, Evertune. Check engines covered, countries/languages, sampling frequency, prompt limits, data export and whether results come from real UI sessions or API calls.
• Classic SEO suites (Sistrix, Searchmetrics - both German; Semrush, Ahrefs) increasingly add AI-answer modules.
• Choose A when: you want dashboards in days and have no developers.
• Check: methodology transparency; vendors differ widely and numbers across tools rarely match.

Path B — Build on cloud APIs

• Query engines programmatically: OpenAI Responses API with web search tool, Google Gemini API with Search grounding, Perplexity Sonar API, Anthropic Claude API with web search tool, Mistral (EU provider) with its web-search connector. API answers can differ from the consumer apps, so treat them as a proxy.
• Analysis: an LLM with structured outputs extracts mentions, citations and claims; embeddings cluster prompts and cited pages; results in a warehouse (BigQuery, Snowflake, Postgres) with a BI tool.
• Orchestration: n8n, Airflow or a cron job; store raw answers for audit.
• Choose B when: you need custom prompt sets, many markets or integration with your own BI.

Path C — Self-host (open source)

• Open-weight LLMs (Llama, Qwen, Mistral families, gpt-oss) served with vLLM or Ollama can do the classification, judging and summarising of collected answers; embeddings via BGE or similar.
• Collecting answers still requires calling the commercial engines or their APIs. Scraping consumer chat UIs usually violates terms of service; prefer official APIs.
• Own-site auditing: Lighthouse, Screaming Frog, Scrapy/Crawl4AI, a robots.txt and schema.org validator. Hardware: HW0-HW1 for the audit; HW2 only if you judge large volumes locally.

Path D — Custom training

Rarely sensible. A fine-tuned small classifier for mention/sentiment labelling is possible for agencies monitoring thousands of brands. Training data: a few thousand human-labelled answers.

State of the art (2026)

• Engines: ChatGPT Search, Google AI Overviews and AI Mode, Gemini, Perplexity, Microsoft Copilot and Claude with web search are the main answer engines; they combine live retrieval with model knowledge and cite a small set of sources.
• Technical access: OpenAI documents separate crawlers (OAI-SearchBot for search inclusion, GPTBot for training, ChatGPT-User for user-triggered fetches), so a site can allow search while opting out of training [3]. Google controls AI use of content through Googlebot access, nosnippet and related directives, and offers Google-Extended for Gemini training and grounding opt-out.
• What tends to be cited: well-structured, factual, up-to-date pages; documentation; third-party reviews, comparison articles, Wikipedia, forums such as Reddit. Earned mentions matter as much as the own site. Structured data (schema.org [4]) helps machines to interpret pages, but engines state no guaranteed ranking effect.
• llms.txt (a proposed markdown index for LLMs [5]) is a community proposal; major engines have not confirmed that they rely on it. Treat as optional and low-cost.
• Limits and failure modes: high run-to-run variance; personalisation and location effects; vendor tools measure different prompt sets; no official engine reports exist (no "Search Console for ChatGPT"), so attribution of traffic is incomplete; hallucinated facts about brands persist even after sources are fixed; manipulation tactics (hidden text, prompt injection into pages, fake review farms) are penalised or filtered and carry legal risk.

Shared resources & combinations

• Same stack (llm-general, web-scraping, text-classifier, workflow-automation): LANG-017 competitive intelligence, KNOW-015 deep-research agents, LANG-013 sentiment analysis.
• Content side: LANG-005 SEO content production, LANG-011 brand-voice rewriting; REC-006 for campaign personalisation.
• Measurement: REC-005 marketing mix modelling and attribution (add AI referral as a channel); KNOW-005 for on-site search and product data quality, REC-001 for shop recommendations.
• Monitoring prompts and judge pipelines are reusable for AI-answer quality checks in other use cases.

Effort, cost & KPIs

• Path A: hours to set up; SaaS subscriptions typically scale with number of prompts, engines and markets; entry plans are affordable for SMBs, enterprise plans cost considerably more.
• Path B: 2-6 weeks for a first pipeline; API costs scale with prompts × repeats × engines × markets; budget for re-runs weekly.
• Path C: 1-3 months incl. ops; only worthwhile for volume or confidentiality needs.
• KPIs: mention rate and citation rate per topic, share of voice versus named competitors, answer accuracy (share of correct brand facts), sentiment, AI-referral sessions and conversions (from analytics referrers), AI crawler hits in server logs, time to correct a wrong claim.

People & change

Management sets the goal and accepts that visibility in AI answers is hard to control. The marketing or communication lead owns the process and the web editors do the work; IT handles the CMS and structured data, and legal or compliance reviews claims about products so that corrected content stays accurate and fair.

Works council: no, because the tool monitors external answers and no employee behaviour. Check only if editor activity would be logged individually; §90(1) no. 3 BetrVG still calls for timely information if work processes change noticeably. A DPO check is a short formality unless prompts contain personal data.

AI literacy: support the editorial team with a short introduction to how AI assistants pick sources and why answers differ between runs (AI Act Art. 4 supports, no guaranteed level). Pilot with one product line or location and a fixed prompt set, then compare against the baseline. Typical concerns are unrealistic promises from agencies and fear of manipulation; address them with honest KPIs (mention and citation rate), no hidden text or fake reviews, and a human sign-off on all published content.

See GUIDE-001, GUIDE-009 and GUIDE-007.

Risks, regulation & compliance

• EU AI Act: the monitoring itself is minimal risk; AI literacy (Art. 4) applies. Content generated with AI and published to inform the public on matters of public interest may need disclosure under Art. 50 unless human-reviewed with editorial responsibility. Not an Art. 5 area, but manipulative techniques targeting vulnerable users would be [6].
• GDPR: prompts and answers about named persons (founders, doctors, craftspeople) are personal data. Right to rectification and erasure can be asserted against engine providers; noyb has filed complaints on chatbot hallucinations about individuals. Do not feed customer data into tracking prompts. Check DPAs with tool vendors and where they process data [7].
• Unfair competition / advertising law: UWG and the Unfair Commercial Practices Directive prohibit fake reviews and undisclosed advertising; the Omnibus rules on reviews apply. Sponsored mentions must be labelled.
• Copyright/TDM: robots.txt and machine-readable opt-outs under Art. 4 DSM Directive and the AI Act GPAI code of practice matter for deciding which crawlers to allow.
• DSA: large search engines must explain recommender parameters; useful for complaints, no entitlement to a ranking.
• Mitigations: official APIs, ground-truth fact sheet, human approval, no deceptive tactics, documented methodology.

Getting started

1. Build a fact sheet (name, offers, prices, locations, differentiators) and 50 realistic customer prompts in your market languages.
2. Check access: robots.txt rules for OAI-SearchBot, Googlebot, PerplexityBot, Bingbot; server logs for AI crawlers.
3. Run a baseline for two weeks (manual or a tool trial) with each prompt repeated several times.
4. Fix basics first: accurate business profiles, clear factual pages, schema.org, fresh dates, FAQs with real answers; then earn mentions in the third-party sources that get cited.
5. Re-measure monthly, track AI-referral traffic in analytics, and report ranges instead of ranks.

Sources

1. Aggarwal et al., "GEO: Generative Engine Optimization" (KDD 2024) - https://arxiv.org/abs/2311.09735
2. Google Search Central - AI features and your website - https://developers.google.com/search/docs/appearance/ai-features
3. OpenAI - Overview of OpenAI crawlers - https://platform.openai.com/docs/bots
4. schema.org - Product and Organization vocabulary - https://schema.org/docs/documents.html
5. llms.txt proposal - https://llmstxt.org/
6. Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj
7. Regulation (EU) 2016/679 (GDPR) - https://eur-lex.europa.eu/eli/reg/2016/679/oj
8. Perplexity - Sonar API docs - https://docs.perplexity.ai/