How GEO works, in four phases

Getting cited by an AI assistant comes down to four things in order: prove whether the engines can reach you, fix what stops them, make the wider web agree on who you are, then measure again next month against the same questions.

Measure. Build a fixed prompt set from how your buyers actually ask, run it through the ngines, and record where you stand today.
Fix. Repair access, structure and markup, and hand over both the files and the guide to deploy them.
Build. Make the entity resolvable, so the web outside your domain agrees on what your brands.
Re-measure. Re-run the identical prompt set every month, against history, so movement is provable rather than claimed.

The order matters. Measuring after the work has started leaves you with no baseline and no way to prove anything moved.

Phase 1: Measure

*What happens:** we build a prompt set from the questions your buyers actually type, freeze it, and run it across ChatGPT, Perplexity, Gemini, Google AI Overviews, Claude and Copilot. Then we test whether each named AI crawler can physically reach your pages.

Freezing the set is the point. A prompt set that changes between runs produces a number that cannot be compared to last month’s number, which is how most AI visibility reporting quietly becomes decorative.

Delivered in this phase:

1. Citation tracking across ChatGPT, Perplexity, Gemini, Google AI Overviews, Claude and Copilot
2. A frozen, re-runnable prompt set, built per market and per language, which you own
3. Mention, citation and self-citation counted separately
4. Share of voice against competitors you name
5. Competitive gap analysis: who is cited in your place, on which page, and why
6. Geo-language matrix: visibility broken out by country and language
7. Source attribution: which of your own pages get cited, and which never do
8. Per-crawler access testing through real fetches as GPTBot, ClaudeBot and PerplexityBot
9. CDN and WAF edge-block detection across Cloudflare, Akamai and platform edges
10. JavaScript-dependency testing: what the page contains before any script runs
11. Void-run exclusion and cold-prompt testing, so a single lucky answer is not mistaken for a position
12. Technical readiness score out of 100, with the breakdown adding up to the number
13. Effective score: readiness adjusted for whether crawlers can actually reach you
14. Negative signal detection: promotional density, keyword stuffing, boilerplate ratio, broken links

Why two scores. Readiness is what your site would be worth if engines could read it. Effective is what it is worth given whether they can. A site can score 68 out of 100 on the work and zero in practice because a WAF answers GPTBot with a 403. We have measured exactly that, on a market leader.

What you get: a report showing where you stand, which engines cite you, which cite a competitor instead, and the specific reason for each.

Phase 2: Fix

What happens: we generate every file and change the audit identified, then hand them to your developers with an implementation guide naming the exact file, the exact location and the verification step. Or we deploy them ourselves. Your call, and the artifacts are yours either way.

Access work:

1. robots.txt rules per named crawler: GPTBot, ChatGPT-User,
2. ClaudeBot, PerplexityBot, Google-Extended, Bingbot, CCBot
3. CDN and WAF rule changes so search-and-cite crawlers get through while scrapers stay out
4. Server-side rendering of content that currently appears only after JavaScript XML sitemap and sitemap-index correction

Machine-readable layer:

1. llms.txt and llms-full.txt
2. agent-card.json for agent-to-agent discovery
3. Clean markdown mirrors of key pages
4. pricing.md, so agents evaluating you on a buyer’s behalf can parse what you charge
5. Open Knowledge Format bundle where it fits the site

Structured data, in JSON-LD:

1. Organization and WebSite
2. Article and BlogPosting
3. FAQPage
4. HowTo
5. Product, Service and Offer
6. LocalBusiness
7. BreadcrumbList and ItemList
8. Review and AggregateRating
9. A stable entity @id shared across every block, so the schema describes one thing rather than several

On-page:

1. Meta titles, descriptions, Open Graph and Twitter cards
2. Canonical tag audit and correction
3. Semantic HTML and heading hierarchy repair
4. Accessibility tree quality, which is how agents navigate a page
5. Core Web Vitals and render performance
6. Internal linking and topic-cluster architecture
7. Visible last-updated dates and named author attribution

Content:

1. Answer-first restructuring into self-contained to word passages
2. Headings rewritten to match how people actually ask
3. Entity-led topic clusters covering the fan-out queries around your parent topic
4. FAQ and conversational sections built from real question phrasing
5. Comparison and alternative pages, the highest-citation content format
6. Definitive guides on your category terms
7. Original research and data studies, the most citable asset a brand can own
8. Research published as a named, versioned, dated index on a canonical URL
9. Dataset schema carrying distribution, licence, measurementTechnique and a citation string
10. A permanent DOI minted per run via Zenodo deposit
11. A pre-formatted “cite this study” block, APA and plain text
12. Third-party amplification per run: trade press, community, direct notification to everyone measured
13. Statistic and citation enrichment of existing pages
14. Multilingual versions for every market your buyers ask in
15. Content decay monitoring and a scheduled refresh cycle

The content work follows the Princeton GEO findings: citing sources, adding statistics and adding quotations were the three strongest levers across roughly 10,000 test queries. Keyword density was the weakest and measurably reduced visibility, so we remove it where previous agencies added it.

Phase 3: Build the entity

What happens: we make the web outside your domain agree on what your brand
is. This is the part site-only work misses, and it is why two companies with
identical technical scores get cited at different rates.

A model needs to resolve your brand to one stable thing. That resolution is built from external identifiers and consistent descriptions, not from your homepage.

1. Wikidata entry creation and maintenance, giving the entity a stable Q-ID
2. Wikipedia accuracy work, where genuine notability exists
3. sameAs links connecting your schema to Wikidata, Wikipedia, LinkedIn and Crunchbase
4. Consistent brand description across every third-party surface
5. Digital PR and mention building in publications your buyers read
6. Review platform profiles: G2, Capterra, TrustRadius
7. Reddit, forum and community presence, through genuine participation
8. Placement in third-party roundups and comparison articles
9. YouTube content for the how-to queries in your category
10. Google Business Profile for local and service entities
11. Merchant Center feed quality for product entities
12. Citation network density: being referenced by sources the models already trust

Phase 4: Re-measure

What happens: the frozen prompt set from Phase 1 runs again every month against the same engines, recorded to a history database. You get the delta, not a fresh snapshot.

1. Monthly re-run of the identical prompt set
2. Month-over-month trend from a persistent history database
3. Which competitor gained or lost the citations you did
4. Sentiment of how engines describe your brand
5. AI referral traffic attribution into GA4
6. Server log analysis of AI crawler behaviour
7. A re-measurement baseline you keep, and can re-run without us

A single screenshot proves nothing, and there is a paper on it. 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 are different the next. Any single reading of citation rate can swing materially, which means a snapshot showing you absent may just be the run where you were absent.

That is why the prompt set is frozen and the cadence is fixed. It is the only way to distinguish work that moved the needle from normal variance.

Reporting the rung, not a single number

Most tools report a citation count and label it “AI visibility”. Citation and recommendation are different outcomes with different causes, so we report them separately. See the visibility ladder.

1. Visibility ladder reporting: retrieved, cited, mentioned and recommended, each reported separately
2. Mention framing tracked rather than counted: recommended, neutral, hedged, or recommended against
3. Recommended-against detection on requirements-heavy queries, with the sources traced
4. Branded search volume tracked as an AI-influence proxy
5. Self-reported attribution captured and wired into reporting
6. Sales call language reviewed for AI-shaped shortlists

Why these six matter more than they look. Similarweb found that 55.9% of AI-influenced traffic arrives through search rather than as an AI referral, against 40.4% for visits with no AI influence. The assistant makes the recommendation, the buyer then types your name into Google, and the visit lands as ordinary branded organic. A client judging this channel by the referral line in GA4 sees the smaller half of the effect. These six rows are how the rest gets reported.

How we work with you

1. A named consultant assigned to your account for its duration, not a shared ticket queue
2. Continuous contact throughout the engagement, not a report and a handover
3. Strategy built per industry, market, target audience, company size and country
4. Scope and price quoted per engagement, agreed in writing before work starts

Why there is no rate card

A person works on your account, so the hours follow the size of the job. A five-page site in one language and a 400-page catalogue across four markets are not the same work, and pricing them the same would mean overcharging one and underserving the other.

What drives your quote is published rather than hidden: page count, number of markets, number of languages, engines tracked, the current technical state established by the free report, whether we implement or hand over to your developers, and how much entity work the brand needs. The free report comes first, so the quote is based on measurement rather than on a guess about your site.

The comparison worth making. AI visibility platforms cost roughly €29 to €500 per month and give you a dashboard. Nothing in that price fixes anything. The trade is a subscription and your own team’s time against a quote and a person who does the work. If a dashboard covers what you need, buy the dashboard.

Payment structures

A quoted fee is the default. It is not the only option, and the right structure
depends on how the business earns.

Structure	Fits
Quoted engagement fee	Most accounts. Scope known, price fixed against it
Percentage of turnover	High-volume sectors that want costs moving with the season rather than sitting as a fixed line. Tour operators usually prefer this and we usually recommend it there
Performance-linked component	Where agreed measurement outcomes can carry part of the fee
Milestone-based	Longer engagements paid across phases
Retainer	Continuous work rather than a defined project

Ask for the structure you want. We will build a payment solution around how your business actually runs, and the numbers go in writing before work starts.

Built per market, not per template. A Polish tour operator, a German B2B manufacturer and a Greek hospitality group need different prompt sets, different languages, different competitors and different engines weighted differently. The method is fixed. Everything inside it is built for the account.

Why the audit runs on infrastructure, not by hand

Most GEO audits are manual, which caps how much can be measured and makes the next run a fresh piece of consulting rather than a comparison.

Ours runs through a containerised pipeline. The practical differences:

Scale. We audited sixteen European tour operators, in four languages, in a single afternoon. Manual methodologies quote six to eight weeks for one enterprise engagement.

Repeatability. The same measurement runs identically next month. Nothing depends on which analyst ran it or what they remembered to check.

Bot-level truth. Access is tested by requesting your pages as each named crawler and reading what comes back, rather than inferring from a robots.txt file. Both failure modes we find most often, a CDN returning 403 and a page that renders to nothing, are invisible to file inspection.

What we will not do

Stating this matters more than another capability line.

We do not cloak. No crawler-only version of a page, ever. Engines detect it and the penalty is removal rather than a ranking dip. On JavaScript-heavy sites somebody always proposes it as a shortcut, and the answer is no.

We do not fabricate mentions. No bought Reddit threads, no seeded reviews. Google’s guidance names inauthentic mentions directly, and a poisoned entity is harder to fix than an absent one.

We do not write separate content for machines. Google calls AI-targeted content variants and chunking-for-AI scaled content abuse. Everything ships written for a reader.

We do not guarantee citations. Nobody controls a model’s output. We guarantee the measurement, the implementation and the baseline.

We do not sell llms.txt as a ranking factor. We ship it. It also sits on around 10.13% of domains with no demonstrated correlation to citation rate, and Google has said publicly it does not use it. You should know that from us rather than from someone else later.

Frequently Asked Questions

How long before I see movement?

Technical fixes get picked up in weeks. Entity and authority work compounds over months. The first re-measurement lands 30 days after the baseline, which is the first point where a change is provable rather than anecdotal.

Do I have to give you access to my site?

No. We can hand every artifact to your developers with an implementation guide naming the exact file, location and verification step. Clients who prefer us to deploy can grant access, and the choice does not change the deliverables.

What if my site is behind Cloudflare or Akamai?

That is one of the two most common reasons a well-built site scores zero. We test for it directly and the fix is a configuration change, not a rebuild. You keep blocking scrapers while search-and-cite crawlers get through, because every crawler identifies itself separately.

Can you work alongside our existing SEO agency?

Yes, and it is usually the right structure. The technical foundations overlap, and the entity, extraction and measurement layers are additive rather than competing.

What do I keep if we stop working together?

The prompt set, the history database, every artifact, and the implementation guide. The baseline is yours to re-run with or without us.