THE CITABILITY ENGINE
Private brief · Pepijn Akkermans · Amsterdam Confidential
Prepared for review Amsterdam · 2026 · v01 Not for distribution
The thesis in one line

Search is being replaced by recommendation. Nobody can measure it yet.

Buyers have started asking AI engines which brand to trust instead of scrolling ten blue links. The engine names two or three companies and the rest vanish. The brands cannot see whether they are named, and they cannot see why. That blind spot is the product.

The bet
A free, deterministic tool that scores any page for how citable it is by AI — reproducible to the decimal, costing almost nothing to run — used as the top of a funnel that feeds two revenue engines beneath it. The honesty is the moat: in a field full of dashboards that fake their numbers, the one that shows its work and names its own limits is the only credible one in the room.

01
What actually changed

The unit of competition stopped being the webpage.

Traditional search ranks pages. AI search retrieves paragraphs, scores each one for credibility, and composites an answer that names a handful of brands. A decade of SEO assets — backlinks, domain authority — barely predict who gets named. The leaderboard has been wiped and almost nobody has noticed.

44%
of buyers now use AI as a primary research source before purchase.McKinsey CMO Survey, 2026 · 1,000+ CMOs
16%
of brands systematically track whether AI names them. The gap between demand and instrumentation is the whole opening.McKinsey CMO Survey, 2026
r=0.18
correlation between domain authority and AI citation. Near-random. The old playbook does not transfer.Wellows, 2026 · n=15,847

A buyer asks ChatGPT for the best brand in a category. It names three. Every other business in that category just lost the sale and will never see the moment it happened.

02
The question a credibility person asks first

"How do you know what people ask the AI — and whether a brand got named?"

This is the question that exposes every fake in this category. Here is the honest answer, stated before anyone has to ask. We treat the limit as the foundation, not the fine print.

What is impossible — and what we do instead
You cannot see real AI prompts.
No one can. OpenAI, Google, Anthropic publish zero real prompt volume. There is no Search Console for ChatGPT. Any tool claiming "what people ask AI about you" is inventing it. We never make that claim.
So what is measurable?
Two real things. One: run a chosen set of buyer questions through the engines yourself and record who got named — a benchmark, not traffic. Two: score a page deterministically against the structural signals that cause citation. Both are reproducible. That reproducibility is the credibility.
Why won't a competitor fake it past you?
They already do — most "AI score" tools send a brand name to a model and let it guess a number. Run it twice, get two answers. Ours fetches the actual page and counts actual signals. Same input, same output, every time. An auditor can verify it.

In a one-year-old field where nobody has a decade of experience, the credible position is not pretended authority. It is: here is exactly what the research shows, here is what we can measure, here is what we cannot. That is the same voice that wins the customers — pointed at the product.

03
The product — split by what costs money

One scanner that is free because it is free to run. One tracker that earns because it costs.

The architecture follows the economics. The viral, rankable top of funnel costs almost nothing per use, so it stays free forever. The part that spends real money per check is the part that charges.

Engine one — the hook
The Citability Scan
Free · unlimited
Paste a URL. The engine fetches the live page and scores it across four structural layers that cause AI citation — fetchability, evidence density, extractability, topical depth — then returns the exact fixes. No model guessing a vibe. Deterministic measurement against a published rubric.
Cost to run≈ €0 / scan
ReproducibleYes — to the point
JobAcquisition + ranking
Engine two — the revenue
The Citation Tracker
Lock a set of buyer questions for a domain. The tracker runs them across the engines, samples each one because AI answers are non-deterministic, and reports citation rate, position, and which competitors got named instead — re-checked weekly. This is the automated version of a manual Monday routine.
Cost to runReal API € / check
Engines automatable3 cleanly, not 4
JobRecurring revenue

Two more limits stated plainly, so they are never a surprise: three engines (ChatGPT via Bing, Perplexity, Google AI) automate with clean public citation access — Claude and Gemini do not, so the automated product tracks three and says so. And a browser-only tool cannot fetch arbitrary sites, so the web scanner runs a thin backend to do the fetching. Neither is a problem; both are designed around.

04
Why the free thing is the business

The scanner is not the product. It is the mouth of the funnel.

A free scan that finds invisible brands is a lead magnet that qualifies itself. A low score is the sales pitch. The same intake splits into two revenue motions sized to the customer — and both are fed by something that costs almost nothing to operate.

FREE CITABILITY SCAN Paste a URL · get a real score ≈ €0 to run · unlimited · the ranking asset a low score qualifies them SPLIT BY SIZE & INTENT small & self-serve · serious & high-touch SELF-SERVE TRACKER Low price · bounded · scales Volume play. Thin margin, near-zero marginal cost. The asset you could sell. DONE-FOR-YOU SERVICE High-ticket · capped clients High margin per client, capacity-limited. Cash engine. Funds everything else. One acquisition channel. Two ways to monetise. The channel is the part that compounds.

Stated honestly for an operator's eye: the high-touch service is a beautiful margin that caps at a few clients — it does not scale in the venture sense, but it funds the build and proves the method publicly. The self-serve tier scales but runs thin. The thing worth backing is the free scanner that feeds both and costs almost nothing to keep running. That is where the leverage lives.

05
The economics, told straight

The whole model lives or dies on one number: cost-to-serve.

This is the part most "cheap AI tool" pitches hand-wave. Here it is in the open. The free engine is genuinely free to run. The paid engine spends real money per check, which is precisely why it is bounded and priced — never "unlimited for pennies."

Operation
Cost to run
Free citability scanpage fetch + deterministic scoring
≈ €0.00
One tracked query, sampled across 3 enginesthe honest, multi-sample version — not a one-shot guess
≈ €0.10–0.25
Card processing on a small chargethe silent killer of sub-€3 pricing — fees alone
≈ 12%+
Done-for-you service, per client / monthcapacity-bound, high margin
€3,500

The design consequence is deliberate: free where it is free, bounded where it bleeds. The self-serve tier is capped in query volume so revenue always clears cost-to-serve and fees. The lesson the market already taught — the cheapest credible competitors floor around €25–29/mo, not because of greed but because fees plus API plus support stack up — is built into the model rather than discovered later.

06
Why this is defensible

The code is copyable in a weekend. The position is not.

A scanner is a few hundred lines. Anyone can clone it. The defensibility is in four things that do not copy.

i.
The tool ranks itself
The single sharpest proof in the category: the AI-visibility tool that AI itself recommends first. We apply the method to our own site until it is the named answer for our own category. The product is the case study.
ii.
Reproducible honesty
Competitors guess scores with a model. Ours are deterministic and auditable. In a field defined by fakery, "run it twice, get the same number" is a positioning weapon no bloated incumbent can match without rebuilding.
iii.
Named, from-zero proof
Real brands, real names, real citation curves starting at zero — published, including the ones that move slowly. No anonymised "a leading brand." Honest named results self-select the buyers who matter.
iv.
A closing window
Roughly an 18–24 month window before LLM training stabilises the new citation hierarchy and intervention effects flatten. The interventions favour challengers over incumbents — first movers lock in citations others cannot later dislodge.
07
Why this window, not later

Demand is here. Instrumentation is not. That gap closes.

Three forces are converging right now. The relative bars below are illustrative of direction and asymmetry — the precise figures are sourced beneath — but the shape is the point: enormous, rising demand against almost no measurement.

Buyer adoption
44%
Brands measuring
16%
Old-SEO predictive power
r .18
Window remaining
~18mo

By 2028, an estimated $750B in US consumer revenue is expected to route through AI search. The brands that start building citation now hold a compounding advantage in two years — the ones that wait inherit a hierarchy already set against them.

Sources: McKinsey CMO Survey 2026 (adoption, measurement); Wellows 2026 (DA correlation); Princeton/KDD 2024 (intervention asymmetry, window); McKinsey 2025 ($750B). Full evidence table available on request.

08
Why this founder, for this problem

Built by someone whose entire discipline is process over outcome.

Pepijn Akkermans. Sport scientist, twelve years racing endurance, founder of a performance-coaching practice built on cited evidence rather than vibes. The same instinct — name the source, measure the same thing every week, refuse to fake the number — is what this category is missing.

The GEO method here is not improvised. It is read from Princeton's KDD study, the GEO Citation Lab corpus, McKinsey, and Wellows — turned into a four-layer protocol with an effect size behind every intervention and a refund clause behind the whole thing. The product is that protocol, made self-serve.

The relevant fit for this round specifically: the one skill that makes the free tool's acquisition model survive — ranking it where others cannot — is the founder's existing craft. The distribution is not a hope. It is the day job.

And the honesty that this audience values is not a tactic he is adopting for the pitch. It is already on the live customer site, in writing: "I have no client logos yet and I would rather say that than fake them." The brief simply points that same standard at the product and the numbers.

≈€0
to run the growth loop
2
revenue engines, one funnel
3
engines automatable, stated
~18mo
window to lock position
What I'm asking

Not for a cheque today. For your read.

You have spent a career on what makes something credible on the internet. That is the exact axis this is built on. So the only question that matters first: does the logic hold, is the honesty real, and is this the right shape for the window we're in?

If you think the idea is right, I'll keep building toward the first scanner and the first named case study — and the next conversation can be about what it would take to go faster.

Pepijn Akkermans Amsterdam · 2026 · Confidential — not for distribution