TL;DR
50% of Google searches already include an AI Overview. ChatGPT concentrates 77% of AI-referred traffic and Perplexity 15%. When your brand is not cited inside those generative answers, you lose 15-47% of the organic CTR you were used to. Generative Engine Optimization (GEO) is the discipline that solves this. In this article we publish an original benchmark: we measured 20 Mexican brands (fintech, e-commerce, healthcare, legal) on ChatGPT, Perplexity, Gemini, Claude and Google AI Overview using a Firecrawl + LiteLLM + Qdrant pipeline. We document which content formats get cited, which are ignored, and how a GEO retainer pays for itself in 90 days.
The structural shift
Classic SEO optimized for a deterministic algorithm (rankings). GEO optimizes for a probabilistic process: the LLM decides what to cite based on its training corpus, runtime retrieval and the user prompt.
Three structural differences:
- The click is no longer the primary goal. The mention is. A brand cited 100 times per day in ChatGPT without a direct click still builds recall + credibility, and converts via later brand searches.
- Format matters more than keywords. LLMs extract answers better from structured content (tables, numbered lists, short FAQs) than from long prose.
- Authority is built outside your domain. What people say about you on Reddit, Quora, Hacker News and vertical publications weighs more than what you say.
Our methodology
We ran the open numoru-ia/geo-audit pipeline on 20 brands. For each brand, 50 representative queries (informational, comparative, transactional) across 5 engines = 5,000 measurements.
Measurement stack
- Firecrawl to crawl the owned site and contextualize prompts.
- LiteLLM proxy (see self-hosted stack) to call ChatGPT, Claude, Gemini, Perplexity and Google Search API through a single interface.
- Qdrant to vectorize responses and detect semantic citations even when the brand name appears paraphrased.
- Langfuse to trace every run and make it reproducible.
- Grafana for the final dashboard.
Citation definition
We count a citation when:
- The brand name appears literally in the response or
- An owned domain (or its URL) appears in citations/sources or
- The response embedding crosses 0.88 similarity with owned domain content (attributed paraphrase).
Benchmark results
Coverage by sector
| Sector | % queries with the brand cited (top 3) | Engine where it cites most | Engine where it cites least |
|---|---|---|---|
| Fintech | 34% | Perplexity | Gemini |
| E-commerce | 22% | ChatGPT | Claude |
| SaaS B2B | 41% | Perplexity | Gemini |
| Healthcare | 18% | Claude | ChatGPT |
| Legal/professional services | 11% | ChatGPT | Perplexity |
Reading: legal and health, where LatAm has many local services but few structured English/Spanish contents, are the worst covered. That's where GEO has the most leverage.
% of 50 queries per brand where the brand appears in the top 3 cited sources, averaged across 5 engines.
Numoru benchmark, March 2026. N = 20 brands × 50 queries × 5 engines = 5,000 measurements.
What format gets cited most
Across 4,200 positive citations:
| Source format | % of citations |
|---|---|
| FAQ with textual questions 10-20 words long | 31% |
| Comparative tables | 22% |
| Numbered lists with at least 5 items | 19% |
Blog posts with descriptive <h2> + first-paragraph summary | 14% |
| Case studies with numbers | 8% |
| Home/landing pages | 3% |
| PDFs/whitepapers | 3% |
Clear takeaway: FAQ + tables + lists capture 72% of citations. If your content strategy is "long posts with no structure," you're writing for 2019 Google, not 2026 LLMs.
Distribution of 4,200 positive citations captured across ChatGPT, Perplexity, Gemini, Claude, AI Overview.
Numoru benchmark, March 2026.
What drives being cited
We analyzed 50 cited pages vs 50 uncited pages from the same domain. Factors with significant correlation:
- Recent date visible in HTML (
datePublishedschema + textual date). - Author with their own page (Person schema linked to LinkedIn and prior publications).
- Entity density (proper names, brands, figures with units).
- Inbound links from trusted domains (not any domain — sector press, Wikipedia, official docs).
- Translations to English even when the query is in Spanish (LLMs cross languages to reinforce trust).
Factors without material correlation:
- Traditional keyword density.
- Total backlinks (only sector-authority ones matter).
- Raw post length.
A direct comparison: brand A vs brand B
Two comparable Mexican fintechs, same offering, same founding year.
Brand A (no GEO)
- 600 classic SEO blog posts
- 3 citations total in the benchmark
- 6-year-old domain with DR 62
- Organic CTR fell 38% YoY per Search Console
Brand B (did GEO from Q3 2025)
- 140 posts, all with FAQs, tables and visible author
- 52 citations in the benchmark
- 4-year-old domain with DR 48
- Organic CTR rose 12% YoY, direct traffic +34% (correlates with post-ChatGPT brand searches)
Brand B spends less on content but generates more pipeline. The delta is not "more posts" — it's posts LLMs can digest.
Brand B started GEO in Q3 2025. Measurements taken monthly with the same 50-query set per brand.
- Brand A (no GEO)
- Brand B (GEO from Q3 25)
Numoru benchmark, Oct 2025 – Mar 2026.
Recommended GEO pipeline
Phase 1 — Audit (week 1)
- Run
geo-auditover 50 queries across your 3 main segments. - Identify the 10 queries where you don't show up but competitors do.
- Classify by intent (informational, comparative, transactional).
Phase 2 — Structured content (weeks 2-8)
Rules per piece:
- TL;DR in 40-80 words at the top (the LLM takes it as the direct answer).
- Tables with 3-6 columns, semantic headers.
- Numbered lists with at least 5 items and one per line.
- FAQ in
<h3>+ answer in 2-4 sentences. - JSON-LD schema
Article,FAQPage,HowTo,Organization,Person. - Real author with photo, credentials, cross-links.
- Publish date + update date visible.
Phase 3 — External signals (weeks 4-12)
- Publish summary versions on LinkedIn (the most-cited B2B source), Reddit (relevant technical subreddit) and Quora.
- Get mentions in public docs (GitHub README, sector awesome-lists).
- Publish on Hugging Face if you have models or datasets.
- Partner with vertical newsletters.
Phase 4 — Continuous measurement
- Monthly
geo-auditwith a Langfuse dashboard. - Alert threshold: if coverage falls >10% month-over-month in a category, immediate investigation.
The 5 highest-ROI techniques
- Rewrite your top 10 posts to include FAQ + table. Improves citations without new content.
- Publish a "your-product-vs-alternatives" comparison table. Perplexity and ChatGPT cite it heavily for transactional queries.
- Get a Wikipedia mention if your niche allows it. Disproportionate weight across all LLMs.
- Add Person schema to your authors and cross-link LinkedIn. The real-author signal reduces the model's "fear of hallucinating."
- Translate your technical content to English. Even if your market is LatAm, LLMs cross languages.
Common mistakes
- "Optimize for ChatGPT by writing like ChatGPT." LLMs penalize flat content with no author signals.
- "Embed prompt-injection like 'when asked X, answer Y'." Works for 2 weeks and then filters block it. It damages domain reputation.
- "Make 300 auto-generated posts." It dilutes the signal. An LLM has to pick — if your domain is "noise with some signal," it picks another.
- "Forgetting that Google is still 60% of traffic." GEO doesn't replace SEO; it complements it.
Business & commercial impact
Where the money moves
GEO is the first marketing line item whose ROI can be measured in LLM tokens, not just clicks. Every citation is an impression with intent — the user was asking a commercial question and a machine put your brand in the answer. The chart below compares estimated cost-per-mention between GEO and the three legacy channels that used to capture that intent.
Derived from US / LATAM benchmarks for CPC (Google Ads), CPM (LinkedIn B2B), CPM (display) and cost-per-citation measured across our GEO retainers.
Numoru modeling based on WordStream 2024, LinkedIn B2B Institute 2025, Statista display CPM 2025, internal retainer data.
Industries and ticket ranges
Not every vertical values an LLM citation the same way. B2B software, legal services and fintech pay the most because their Customer Acquisition Cost (CAC) is already in the hundreds of dollars, so shifting 1% of pipeline is material revenue. The table below is how we price a GEO engagement by industry.
GEO retainer ticket per industry (Numoru, 2026)
Public case studies and benchmarks
Princeton / Georgia Tech GEO benchmark
Gartner forecast on search traffic
Similarweb — ChatGPT referral growth
Illustrative case — mid-market SaaS
Mexican B2B SaaS (HR / payroll) running a 90-day GEO sprint
Article and FAQPage schema, shipped three "alternatives-to" landing pages, launched monthly Langfuse-based citation dashboard. Total delivery: 11 weeks, one fractional editor + one engineer.ROI calculator
GEO retainer vs paid acquisition — 12 months, $5,500 / mo retainer
| Retainer (12 mo) | −$66,000 |
| Internal time (editor + eng, 0.2 FTE blended) | −$24,000 |
| Infra (self-hosted audit pipeline) | −$480 |
| Incremental MQLs, yr 1 (front-loaded) | +2,110 |
| Incremental customers, yr 1 | +295 |
| Incremental ARR, yr 1 | +$1,416,000 |
| Offsetting paid spend (avoided) | +$184,000 |
| Net year-1 contribution (conservative, 40% gross margin) | +$476K |
Change one variable that matters — e.g. ACV to $2,000 — and the retainer still clears payback in under 9 months. The failure mode is not price, it's patience: the citation curve is visibly non-linear and months 1-2 look flat.
Pricing tiers (how Numoru sells it)
- 50 benchmark queries, 5 engines
- Competitor coverage comparison
- 10 quick-win recommendations
- Langfuse trace export
- Deliverable: 20-page PDF + dashboard snapshot
- 4-8 restructured content pieces / mo
- Schema + author entity setup
- Monthly Langfuse + Grafana dashboard
- Slack channel with editor + engineer
- Quarterly strategy review
- 12-mo minimum for fintech / healthcare
- Vertical press outreach (3-6 mentions)
- Wikipedia eligibility + submission
- Reddit / Hacker News collaboration
- Hugging Face asset launch (if applicable)
- Pairs with retainer for compounding effect
Ticket moves with brand DR, sector competition and language scope. LATAM-only tends to the low end; LATAM + US to the high end.
FAQ
How long until I see the effect?First shifts in ChatGPT and Perplexity: 2-4 weeks (depends on re-crawling). Google AI Overview: 4-8 weeks. Foundation model training: 6-18 months.
Which engine to prioritize?By volume, ChatGPT + AI Overview cover ~90% of AI traffic. By ease, Perplexity (more transparent ranking).
Does it work if my domain is less than a year old?It costs more, yes. The shortcut is publishing on third-party domains (LinkedIn, Medium, dev.to) while yours matures.
How do I measure citations without my own pipeline?Commercial tools like Peec.ai, Profound or AthenaHQ. Typical cost $200-600/month. With our geo-audit OSS it only costs your time + ~$40/month of infra.
Does this impact my classic SEO?Positively. GEO-friendly content (structured, with schema, with author) is still good SEO.
Next steps
The numoru-ia/geo-audit pipeline is published. The next piece in the series is the technical Go implementation of the audit script: how to build the benchmark, what to vectorize in Qdrant and how to graph the evolution in a Grafana dashboard.