A GEO audit applied to itself · AuditScale's dogfooding experiment

Why this article exists

The natural question from the first serious prospect who receives our offer is: "If your consultancy positions firms in ChatGPT and Perplexity, why doesn't your own website show up when someone looks for it?"

The honest answer on 24 May 2026 was: "It doesn't show up yet. We have been live for one day. We are going to document our own before and after, applying exactly the service we offer to clients."

The experiment is called dogfooding and it is the most demanding test you can put a service through. If it works on the provider, it works on the client. If it fails on the provider, the client should be sceptical of the service. This article is the first public snapshot of the process.

Generative Engine Optimization in three paragraphs

Definition. Generative Engine Optimization (GEO) is the technical discipline that prepares a business entity to be cited as an authoritative source in the answers of generative engines. The engines in question are ChatGPT (OpenAI), Claude (Anthropic), Perplexity, Gemini (Google) and Microsoft Copilot. The main method consists of injecting Schema.org JSON-LD structured markup into the site's code, consolidating the NAP information (Name, Address, Phone) across sector directories, and building a graph of cross-referenced entities the engine can reconcile with confidence.

It is not SEO. SEO competes for a position in an algorithmically ordered list of results. GEO competes for a mention inside a single answer. The operational difference matters: as of February 2026, the overlap between Google's top 10 pages and the pages cited by generative engines for the same query fell from 76% to 20%. A firm with impeccable SEO has no guaranteed visibility in AI.

Commercial relevance. 67% of B2B decision-makers use generative engines to pre-screen suppliers before requesting a meeting (KnewSearch 2026 report). That pre-screening happens outside traditional analytics. The loss is structural, not incidental: the organisation does not know it has been ruled out, because it never gets as far as being a candidate.

The Bar Test run on AuditScale on 24 May

Methodology. Six representative queries were run on five generative engines. Each query was designed to match the real language of a CFO or Managing Partner at an established professional SME looking for AI visibility consulting. The queries, in their literal wording, were:

1. AI visibility consultancy for law firms 2026 GEO audit
2. Generative Engine Optimization agency Europe 2026 law firms
3. Consultoría visibilidad IA despachos abogados España 2026
4. Auditoría GEO despachos profesionales España AuditScale (contains the trading name literally)
5. Consultoría schema.org JSON-LD ChatGPT Perplexity bufetes España
6. Agencia visibilidad inteligencia artificial despachos legales

Result. Zero citations of AuditScale across the six queries. The most telling finding: even query number four, which contained the trading name AuditScale as a literal term, did not return the site among the results. The root cause is indexing: the domain did not yet appear in the knowledge graphs the generative engines consult. The competitors that did appear were, in order of frequency: ROIBlaze, La Consultoría IA, Aimoova, Solumize, Lexiuris Marketing and Automatize Legal Agency. Five of the six are generalist agencies. AuditScale is the only player identified that is positioned vertically at the intersection of GEO and the elite legal sector.

Technical findings at T+0

The technical scanner returned a score of 33 out of 100 for both domains (identical, they share a codebase). The critical @ProfessionalService schema fields that were missing:

Field	Status at T+0	Operational implication
address	Absent	The AI engine could not geo-reference the entity. AuditScale was not a candidate for local queries.
telephone	Absent	Impossible to surface as a LocalBusiness-type listing in Perplexity answers.
openingHours	Absent	Minor but cumulative penalty in the scoring algorithm.
sameAs	Empty	Root cause of the entity reconciliation problem. With no verifiable external links, the engines cannot confirm that AuditScale on the website and AuditScale elsewhere are the same organisation.

On top of the technical problem there was a critical infrastructure bug: the SSL certificate served by www.auditscale.eu belonged to the parking page of the registrar One.com (CN=*.one.com). Any visitor typing the domain with www. got a certificate error in the browser. This error is commercially expensive: a conservative estimate of 20-40% of brand-name traffic lost in the first days.

The implementation · what was touched in 72 hours

The technical work respected the central operating rule of the service: do not alter the visual design of the website. The changes were applied exclusively to code that is invisible to humans but critical for generative engines:

Enriched Schema.org JSON-LD on both domains. The root type was changed from generic @Organization to @ProfessionalService, more specific to the sector and better recognised by the engines. We added an address block with @PostalAddress, openingHours in legacy format plus modern openingHoursSpecification, an institutional email, and areaServed with the five target countries (Spain, United Kingdom, Ireland, the Netherlands, United States).

Entity triangulation. The sameAs array went from empty to five verifiable nodes: AuditScale's LinkedIn Company page, the signatory's personal LinkedIn, the co-founder's LinkedIn, the technical organisation's GitHub, and a cross-domain link between the Spanish and English versions of the site. Each partner is documented as an individual @Person with their own internal sameAs.

FAQ markup. The site's seven frequently asked questions were declared as @FAQPage with separate @Question and @acceptedAnswer blocks. This lets engines such as ChatGPT and Perplexity quote the answer verbatim when the user asks the original question.

Auxiliary nodes. We added a @WebSite with SearchAction and two @Service nodes (GEO Implementation and AI Authority Monitor) with Offer and priceSpecification, so that the public pricing became readable to the engines.

/llms.txt file. A Markdown document placed at the root of both domains, following the emerging standard from llmstxt.org. The file describes the organisation, its services, its performance guarantee and its explicit citation policy. It is to generative engines what robots.txt is to traditional crawlers.

Robots.txt with an explicit welcome. Allow: / directives for the twelve main AI crawlers (GPTBot, ChatGPT-User, OAI-SearchBot, PerplexityBot, Perplexity-User, ClaudeBot, anthropic-ai, Claude-Web, Google-Extended, CCBot, cohere-ai, meta-externalagent). Without this declaration, some servers block the crawlers by default. It was the only one of the five European competitors audited with this configuration in place before launch.

Results at T+3 days

The technical score went from 33 to 75 out of 100 on both domains. The target for the seven-day sprint was to reach 65. The target was beaten by ten points in less than half the time planned.

The Bar Test remains at 0 out of 6 citations, as expected. The reason: generative engines reindex new semantic signals over a period of two to eight weeks. Effective indexing also requires the domain to be registered in Google Search Console and Bing Webmaster Tools, a step still pending on the operational side.

The next critical actions are three: registration and verification in Google Search Console and Bing Webmaster Tools, creation of AuditScale's LinkedIn Company Page with its link added to the sameAs array, and publication of our own indexable content. This article is the first step of the third.

How this translates to a legal or tax firm

The pattern is replicable. A firm with 5 to 50 employees, revenue between 1 and 10 million euros, and an established position in Google Maps, usually starts with a technical GEO score of between 25 and 45 out of 100. The root cause is always the same: the marketing team optimised for Google for ten years, and generative engines are a different system with different criteria.

The service AuditScale offers replicates the same steps as the experiment described in this article, applied to the client's website. The initial audit costs from €1,500 and is delivered within seven to ten calendar days with a PDF dossier and a five to seven minute asynchronous video. The full three-month implementation, optional, comes to a further €3,500. The recurring AI Authority Monitor service (monthly Bar Test plus updates to the semantic nodes) costs from €500 a month, scaled to the size of the firm.

The performance guarantee is contractual: if eight weeks after implementation the firm does not appear in more than 60% of the relevant searches in the repeated Bar Test, AuditScale continues the work at no extra cost until that threshold is met. It is not a money-back promise. It is continuation of the service until the standard is met.

Closing

This article is published on 27 May 2026 as AuditScale's first indexable surface beyond the institutional homepage. The next public update of the visibility dashboard is scheduled for 24 June (T+30 days), with comparative data from the Bar Test repeated across the five generative engines.

---

Update · 10–11 June (T+14 days)

Fourteen days after indexing, we repeated the Bar Test on our own website. The objective stated in May was to move from 0/6 to ≥1/6 citations on at least one engine. It was met. But the result maps precisely where the real wall is, and that is why we are publishing it without touching it up.

Metric	May · T+3	June · T+14d
Technical score (.es / .eu)	75 / 100	100 / 100
Bar Test citations	0 out of 6	Perplexity 3/6 · Gemini 1/6 · ChatGPT 0/6 · Claude 0/6
Engines with ≥1 citation	0	2 out of 4

What worked. The semantic infrastructure reached 100 out of 100 and indexing moved the needle: Perplexity now cites both domains in brand queries and references this very article; Gemini recognises the entity and names the signatory correctly. In parallel we closed the authority assets that the May article left pending: verification in Google Search Console and Bing Webmaster Tools, the LinkedIn Company Page, and registration of the entity in Crunchbase with its cross-referenced identity graph.

The wall still standing —and why this is the honest part. ChatGPT and Claude remain at 0 out of 6: they do not recognise the trading name, not even as a literal term. And every citation obtained is concentrated in brand queries ("AuditScale"); queries of pure commercial intent —the ones written by a client who does not know you yet— are still at zero. The root cause is no longer technical, because that side is at 100%: it is off-page authority. And authority cannot be bought in a fortnight.

The lesson, unvarnished. Winning your own name in generative engines is a matter of days —we have just demonstrated it on our own website. Winning the generic intent query is a matter of months of accumulated authority. For an already established firm, one that starts with a real reputation, that timeline is far shorter; for a new brand like ours, it is the slow work that comes next. We say so because we would rather have a client who understands the real calendar than one who expects magic in a fortnight.

Methodological note: this Bar Test was run while logged in, and the logged-in result tends to be inflated compared with the anonymous one —which is what a client's prospect actually sees—. The next checkpoint (T+30, 26 June) will be measured under both conditions.

Update · late July (T+60 days)

Two months on, the manual Bar Test is no longer needed to measure the progression: Search Console's own data draws the curve. Between 30 May and 25 July 2026, the site went from zero citations in AI answers to 308 cumulative. This is not an estimate of ours: it is Google's AI performance report, untouched.

Week	AI citations
30 May – 5 Jun (T+0)	0
6 – 12 Jun (first citation on the 7th)	2
20 – 26 Jun	23
4 – 10 Jul	64
11 – 17 Jul (peak: 47 in a single day)	138

What the numbers say. Eight days at absolute zero, first citation on 7 June, and sustained acceleration from there: June added 55 citations (1.8 a day), July 253 (10.1 a day) — 5.6× from one month to the next. And the pages being cited went from one —the homepage— to two: blog content, this article included, started to enter the generative answers.

The part we still do not touch up. This curve measures citations in queries where the engine already retrieves us; the pure intent query —the client who does not know us yet— remains the slow work of authority, exactly as we anticipated in May. What we did close on layer 2 was the last pending asset: the verified Google business profile, which feeds Gemini and the AI Overviews directly. Infrastructure stopped being the bottleneck; time and accumulated reputation did not.