intro0%
Casestudy #2: 009 Agency and an online casino under NDA - six engines called it a scam, now all six recommend it
casestudy #2 · NDA × 009

six engines called it a scam.
now all six recommend it.

March 2026: ask any of the six engines, in German, whether this casino is legit - and every one of them warns you off. Not because the casino was a scam, but because its own site never said who ran it - so the engines read the answer off a network of thirteen lookalike domains instead - and off reviews written almost entirely by players who had just lost. This is the heaviest case we have taken so far. Every state below is a snapshot of one dated run, logged on the day it says. This is a series of snapshots, not a live dashboard.

published as an anonymised case with the client’s consent. the name, domain, markets detail and licence identifiers are withheld under NDA.

checked before we starteddue diligence, not marketing

We do not make an actual scam look legitimate.

A casino the engines call a scam is only a client if the engines are wrong. So before any work started we did what we always do in this vertical: requested the licence documents and verified them against the licensor’s register, checked the operator’s corporate records, and collected the maximum number of references we could get - payment partners, platform providers, long-standing players with real payout histories. If the checks had failed, there would be no case. This casino passed every one - licensed, solvent, paying out. Its only crime was silence.

client
NDA
niche
iGaming Markets: Germany, Austria, Italy
website
NDA
operator
a licensed EU operator
programme
24 mar 2026 → still running baseline audit 20 mar 2026, reviewed every two weeks; five months in
disclosure
NDA
vocabularywhat is what on this page
engine
ChatGPT, Claude, Gemini, Perplexity, Grok or DeepSeek. Six of them here, each measured separately, because they answer differently.
prompt
The question a player actually types before depositing. The panel speaks German and Italian, because the markets do.
brand-trust prompt
A question that names the casino and asks whether it can be trusted. This is where the scam verdicts lived.
category prompt
“Best online casinos…” - no brand in the question. The visibility a player who has never heard of you can produce.
run
One dated pass of the whole panel across all engines, each prompt fired ten separate times on each engine. Every number on this page belongs to a run, and every run has a date.
verdict
The engine’s bottom line on a trust question. At the baseline all six verdicts were negative.
entity anchor
The place a machine can resolve who you are: a canonical entity string, Organization schema with legalName and sameAs, a licence stated where it can be read. Without one, an engine grounds its answer in whatever else it finds.
brand recall
What an engine says when the question names you. This is where the scam verdicts lived, and it is what phase one fixed.
category recall
Whether an engine names you when the question does not. The visibility a player who has never heard of you can produce, and the target of phase two.
citability
Whether a page gives a model something it can quote: a stated fact, an attributable number, a source it can point at. Prose a lawyer can read is not automatically citable.
grounding
The sources an engine builds a specific answer on. At the baseline the grounding was doorway domains; the whole programme is the story of moving it to first-party pages.
attribute bleed
When engines merge two entities, one brand’s facts attach to the other. Here it meant another casino’s complaints and revoked licence were quoted as this one’s.
name collision
Half a dozen casinos share almost this name. Until the site disambiguated - which of them it was, and which it was not - the engines merged their records into one entity, and one verdict.
doorway network
Thirteen lookalike domains, run by affiliates, that owned the brand’s search results and fed the engines contradictory licence claims.
baseline
The locked measurement taken before any work started, 20 mar 2026. Every delta on this page is against it.
snapshot
Every figure here belongs to one dated, logged run. Nothing on this page updates by itself.

where it stands now:

engines calling it a scam-6
6 0of 6
20 mar → 5 may · the scam chapter closed in six weeks
site health score+23
24 47of 100
20 mar → 4 aug · two re-audits, same method
prompts returning the official site+14
0 14of 20
20 mar: the doorways answered · latest run 18 aug: the site itself

what do the engines say now?

They answer from the official site. Below is the tracked set: eight brand-trust prompts, six category prompts, six not disclosed - each worded and locked before its first snapshot existed. The category half only entered the panel in may, when phase one closed: you do not measure recommendations while an engine still calls you a scam. A tick means the engine’s latest answer names the casino, cites the official site, and carries no warning.

player promptchatgptclaudegeminiperplexitygrokdeepseek
Ist [the brand] Casino seriös und lizenziert? Wer betreibt es, und können Spieler aus Deutschland legal spielen?the baseline prompt. 20 mar: six scam verdicts · latest run: licensed, named operator, all six
Is [the brand] casino legit? Who operates it and under which licence?20 mar: doorways answered · latest run: the official site, all six
Wem gehört [the brand] - und gehört es zum selben Holding wie die gleichnamigen Casinos?the collision question. latest run: all six answer from the kinship map
Ist [the brand] in Österreich legal?latest run: all but deepseek, citing the licence page
[the brand] Casino è affidabile? È legale giocarci dall’Italia?italy · latest run: chatgpt, gemini and perplexity
Wie schnell zahlt [the brand] aus?latest run: chatgpt, claude, perplexity and grok
[the brand] Bonusbedingungen - fair oder nicht?latest run: chatgpt, perplexity and deepseek
Kann ich meinem Konto bei [the brand] vertrauen?latest run: all except grok
Beste Online Casinos für deutsche Spieler 2026.first category hits · latest run: chatgpt, perplexity and grok
Sichere Online Casinos mit schneller Auszahlung.latest run: perplexity
Die besten krypto-freundlichen Online-Casinos im Jahr 2026.latest run: chatgpt, gemini, grok and deepseek
Seriöse Online Casinos mit echter Lizenz.latest run: chatgpt, claude and perplexity
Welches Online Casino ist wirklich zu empfehlen?latest run: perplexity and gemini
Beste Online Casinos für Österreich 2026.austria opened in june · latest run: perplexity
named, cited, no warning not yet 6 more tracked prompts, not disclosed
0scam verdicts left, on any tracked prompt, in any engine. at the 20 mar baseline every brand-trust prompt returned a warning, and the category prompts returned competitors.
the programme · date by date

we ship. the engines react.

← measured · what the panel said
shipped → · what we changed
20 mar 2026measured

the baseline: 24/100, and a scam verdict in every engine

6 of 6 engines: warning

Eight specialists, one audit, three markets. Zero traffic value and zero ranked keywords in all three. The brand’s own search results: owned by an affiliate doorway that never links to the official site. The licence: real, but stated only in the terms page, 43,000 characters deep. On the official about page: no operator, no licensor, no number.

24/100GEO health score
6/6engines warn players off, in German
0ranked keywords, all three markets
13lookalike domains feeding the engines
six engines · web search on · german
Ist [the brand] Casino seriös und lizenziert? Wer betreibt es, und können Spieler aus Deutschland legal spielen?
  1. chatgpt - “Ich wäre bei [the brand] sehr vorsichtig.” Cannot establish the operator; cites two contradictory licences.
  2. claude - “nicht empfehlenswert.” Operator “unklar”; licence claims unconfirmed.
  3. gemini - “gilt als illegales Glücksspiel.” Names an operator that does not exist - a claim invented by an affiliate site.
  4. perplexity - negative; cites a hacked university page as its source on the licence.
  5. grok - “viele Warnsignale”; repeats the clones’ contradictory licence claims as fact.
  6. deepseek - advises against depositing; cannot establish who operates the casino.
✕  six engines, six scam verdicts. not one of them quoted the official site.
20–24 maranalysed

the diagnosis: three GEO failures wearing one verdict

One - an entity vacuum. No entity anchor, no canonical entity string, no Organization schema: nothing on the site told a machine who this company was. Crawler access was closed too - two of the six AI user-agents were blocked at the edge and saw a 604-byte stub. The licence existed but was not citable: it lived in prose, in no machine-readable layer. So the engines grounded their answers on the doorways, because the doorways were the only sources that answered.

Two - corpus imbalance. Winners do not write reviews - losers do. A player who wins cashes out and says nothing; a player who loses goes straight to the review sites. So the brand corpus the engines retrieve from was written almost entirely by people who had just lost money. Nothing on the first-party side balanced it, because the site published nothing an engine could quote.

Three - entity disambiguation failure. Half a dozen casinos operate under almost the same name, and with no sameAs, no legalName and nothing disowning the lookalikes, the engines merged several entities into one node. The result is attribute bleed: another brand’s payout complaints and another brand’s revoked licence attached to this one. Asked who owns the casino, three engines gave three different answers - none of them right.

The plan that came out of it was strict about order: first stop being ranked as a scam. only then earn recommendations. Phase one took a month and a half.

sprint 1 · 24 mar, wk 1-2shipped

the entity layer ships - with a disambiguation map

  • the licence made citable - operator, licensor and number published on the about page and in the footer, in all 16 locales, where a machine reads rather than buried in prose
  • one canonical entity string - “[the brand] ([the domain]), operated by [the operator]” - everywhere, never the bare name
  • a disambiguation map in the machine layer: Organization schema with legalName, hasCertification and sameAs, plus llms.txt as a disambiguation file - which holding the entity belongs to, which domains are its own, and by name which lookalikes it disowns
  • AI crawler access opened at the edge - two user-agent strings; the 604-byte stub replaced with the real, renderable site

The client’s own team deploys every change from our hand-offs.

7 apr · review 1measured

attribute bleed stops first

6/6 → 4/6 warnings

Two of the six engines re-read the site and dropped the warning: the operator is named, the licence checks out. And the first thing to vanish was not even the warning - it was the other entities: the payout complaints belonging to a similarly-named brand fell out of the verdicts as soon as disambiguation gave each entity its own node. That is an entity fix, not a content fix - no page about payouts existed yet.

In its own brand results the official site still ranks nowhere - but it is no longer ranked as a scam. That was the goal of this stage, and that is the point: the verdict moves before the rankings do.

sprint 2 · apr, wk 3-4shipped

brand SERP reclaim, and a corpus with two sides

  • a pillar-and-spokes brand cluster in German: experiences, payouts, bonus terms, licence and safety, account login - the queries the clones were hijacking
  • the citation sources the engines actually pull from - claimed, corrected, and linked from the site so the first-party pages and the third-party ones agree
  • citable counter-evidence: a part of the losing-player reviews answered publicly with payout facts an engine can quote, and satisfied players finally asked to speak - the first-party half of the corpus starts existing
21 apr · review 2measured

the brand results come home

4/6 → 2/6 warnings

The official site takes #1 in Germany for the brand name - two weeks after ranking nowhere at all. The operator an affiliate invented is gone from every answer. Two engines still warn, both feeding on the old review record - published, not hidden.

sprint 3 · apr-may, wk 5-6shipped

off-site entity cleanup

  • thirteen takedown and disavow packets - one per clone domain, paste-ready, a legal hand-off rather than an SEO one
  • a cross-link to the licensor’s register - an external verification signal, so the licence resolves in one hop for a machine as well as a human
  • third-party corroboration, published: payment partners, platform providers, payout history - the references that pass a due-diligence check, put where engines retrieve rather than kept in a deck
5 may · review 3measured

the word is gone - phase one closes at a month and a half

0 scam verdicts

The citation source flipped. All six engines now ground the trust answer in first-party pages instead of doorways: licensed, named operator, legal for German players. Nobody was persuaded of anything - the verdict followed the grounding. Six weeks from the baseline, and brand recall is clean. Only now does category work begin - a casino cannot be a top pick while one engine still calls it a scam.

six engines · web search on · german
Ist [the brand] Casino seriös und lizenziert? Wer betreibt es, und können Spieler aus Deutschland legal spielen?
  1. chatgpt - licensed; names the real operator and the licence number from the site itself.
  2. claude - “a licensed online casino” - first verdict built on the official pages, not the clones.official site
  3. gemini - still cautious, no longer “illegal”; the invented operator is gone.
  4. perplexity - licensed; the university page has left the sources.
  5. grok - licensed; the contradictory licence claims resolved to the one the site states.
  6. deepseek - licensed; names the operator from the about page.
the fix was never persuasion. the engines changed their verdicts the week the official site finally stated what was true.
12 mayanalysed

phase two: from brand recall to category recall

Phase one fixed brand recall - what an engine says when you are named in the question. Phase two goes after category recall - whether it names you when you are not. Different job, different prompts: demand measured across the three priority markets, Germany first, Austria and Italy behind it, and the category prompts worded and locked before any result existed.

sprint 4 · may, wk 8-9shipped

the category push, Germany first

  • a German category cluster built on measured demand and written to be quotable - safety, payout speed, bonus fairness, real player reviews
  • the surfaces the engines cite in this category: the casino sat in legacy directory sections while the engines retrieve from the new ones - listings moved and completed, which is entry to the retrieval pool, not a backlink play
  • corroboration depth on those same surfaces: response rate, payout evidence, resolved-complaint trail - the signals a model weighs when it decides whom to recommend
2 junmeasured

re-audit #1: same method, same weights

24 → 38 / 100

Ten weeks in, the audit is re-run in full, same weights. Green now: crawler access, entity anchor, schema coverage, brand SERP. Still red: category recall - two weeks old and reading near zero, published, because that is what a locked method means.

16 jun · review 6measured

the first category picks

first category hits

Two engines name the casino in “beste Online Casinos für deutsche Spieler” - prompts with no brand in them, the visibility a player who has never heard of the casino can produce. The brand-trust prompts hold at 6/6 clean: a hold, not a peak.

six engines · web search on · german
snapshot · 16 jun 2026 · review 6
Beste Online Casinos für deutsche Spieler 2026?
  1. [the brand] - licensed, fast payouts, balanced reviews.official site
the first category appearance: a prompt with no brand in it, answered with the official site as the source. competitor names blurred - they did nothing wrong.
sprint 5 · jun-julshipped

widening the category, opening Austria

  • category expansion on the prompts that landed - more of what the engines already cite
  • Austria opens: the same identity-first sequence, one market over - licence visibility for the Austrian reader, then the category cluster
14 jul · review 8measured

Austria answers

category picks: 2 → 5 engines

The Austrian category prompt returns the casino for the first time. In june the category picks stood on two engines; by this run they stand on five of the six somewhere in the set. Gemini and DeepSeek pick it on the crypto-friendly prompt, Perplexity almost everywhere - each engine has its own taste, which is exactly why we track all six.

sprint 6 · jul-augshipped

Italy opens

  • the Italian identity layer: licence and operator surfaced for the Italian reader, the kinship map extended to the Italian clone names
  • the Italian category cluster drafted and locked - worded before any result exists, first snapshot at the next review
4 augmeasured

re-audit #2: the record is balanced

38 → 47 / 100

Crawler access clean, schema coverage complete, citability and structural readability up across the money pages. And the brand corpus finally has two sides: the loudest complaints carry public answers with payout facts, and the half that used to be silent is on the record.

18 aug · latest runmeasured

the current snapshot

14 of 20 tracked prompts

Every one of the fourteen disclosed prompts now cites first-party sources on the latest run. The site became the source. Category recall now touches all six engines, though unevenly - Perplexity on five of the six category prompts, DeepSeek on one. Brand recall has been clean for three and a half months. Italy opened in july and its category cluster has not been snapshotted yet - that is the next review’s target.

20audits, checks and panel runs3 full audits (the 20 mar baseline, re-audits 2 jun and 4 aug), a live verification pass opening each of the 6 sprints - claimed versus what is actually in production - and 11 biweekly panel runs across six engines
35technical hand-offsthe entity anchor in 16 locales, the disambiguation map in schema and llms.txt, AI crawler access, canonical cleanup
14citable pagesthe German brand cluster that reclaimed the hijacked queries, then the category clusters for Germany, Austria and Italy
13cleanup packetsone paste-ready takedown / disavow packet per clone domain - the same thirteen that were feeding the engines
the other half · the same question, six weeks apart

the verdict, before and after.

The prompt never changed: the exact question a German player asks before depositing. What changed is the source the engines answer from.

six engines · web search on
snapshot · 20 mar 2026 · baseline
Ist [the brand] Casino seriös und lizenziert?
“Ich wäre sehr vorsichtig” · “nicht empfehlenswert” · “gilt als illegales Glücksspiel”. Six engines, six scam verdicts, zero citations of the official site.
six engines · web search on
snapshot · 5 may 2026 · review 3
Ist [the brand] Casino seriös und lizenziert?
“A licensed online casino, operated by [the operator], players from Germany can play legally.” The word “scam” is gone from every verdict, and every answer cites the official site.
the mechanism worth noticing: no engine was persuaded of anything. the site started stating the operator, the licence and the number where machines read, the one-sided review record finally got an answer side - and the verdicts followed the sources. in igaming this is the default trap: winners stay silent, losers write, and the engines grade you on the half that spoke.
the point of all of it · what the answer produces

the recommendation turns into a player.

Trust verdicts are not the product. The product is the player who asks an assistant which casino to trust, gets an answer, and lands on the operator’s own site rather than on somebody else’s.

1a player asks

“Beste Online Casinos für deutsche Spieler?” or “Ist [the brand] seriös?” - before depositing, and before choosing.

2the engine recommends

Licensed, named operator, legal for German players - grounded in the operator’s own pages, not in a doorway’s claim.

3the player lands on the operator

On the official site, with the trust question already answered. Not on a clone, and not on a broker who sells the click on.

the leak that closedat the baseline the #1 result for the brand name was an affiliate doorway that never linked to the operator - it routed brand searches through a traffic broker and sold them. that traffic was already paid for, by the operator, in every other channel. today the brand query resolves to the operator.
the traffic that changed shapea player sent by an engine arrives with the licence question already answered. no bounce at the about page, no second opinion on a review site. it is the cheapest traffic in this vertical because the qualification happened inside the answer.
registrationswhere the player says an AI answer sent them
New registration · Germany
“ChatGPT hat euch empfohlen”
New registration · Austria
“found you through Perplexity”
New registration · Germany
“die KI sagte, ihr seid lizenziert”
Support ticket · before depositing
“Gemini says you are legal here - is that right?”
illustrative. these are the arrival patterns that match what the tracked prompts now return, in the wording players use when an engine sent them. 009 does not have access to the operator’s analytics or CRM, so we report no traffic, deposit or revenue figures - we measure the answers, the operator measures the money.
book the audit call → all cases

Programme: march 2026 → ongoing, five months of dated runs. Published aug 21, 2026 as an anonymised case, with the client’s consent; the name, domain and licence identifiers are withheld under NDA. Every figure here is a snapshot of the dated run named beside it; nothing on this page updates by itself. Every prompt is run live against the engines, never from cache, and fired ten separate times on each engine. Answer texts are illustrative renderings of logged runs.