You cannot manage what you cannot see.
Every marketing team in 2026 has a Google Analytics dashboard. Every team has ad performance dashboards. Most have some form of SEO rank tracker. Almost none, even at sophisticated brands, have a baseline on how they are described, cited, or omitted inside the AI systems their customers use daily.
This is not a minor gap. It is the largest unmeasured source of commercial influence in modern discovery. A prospect asking Perplexity "Which is the best AI growth agency in Singapore?" receives a sentence. That sentence, whether it mentions you or does not, shapes the next 24 hours of their research, their shortlist, and eventually their brief. And in most organisations today, there is no system of record for it.
An AI visibility audit closes that gap. It produces an objective baseline, how LLMs describe you, cite you, position you against competitors, and where the gaps are, that becomes the starting point for every subsequent growth decision.
Prompt visibility, do you show up at all?
The first layer is the most visceral. We run 40–80 commercially relevant prompts, the questions real prospects actually type, across the four LLMs that matter (ChatGPT, Gemini, Perplexity, Google AI Overviews) and score every response for the presence of your brand and your direct competitors.
We measure three things per prompt:
- Mention. Are you named at all? If not, you do not exist in that query's consideration set.
- Position. Where in the answer are you mentioned? First named brands carry 2–3x the weight of later mentions in user recall.
- Framing. What sentence is constructed around your brand? "A specialist in…" is very different from "Another option is…", and both are different from "…though users have reported concerns about…".
The output is a prompt-by-prompt table showing exactly where you appear, where you are absent, and where competitors have captured the narrative. For most brands, this alone is revelatory, they discover that on 60–80% of their highest-commercial-intent queries, they are simply not in the answer.
Entity legibility, can an LLM describe your brand accurately?
We prompt each LLM directly: "Describe [your brand] in 150 words. What do they do, who do they serve, where are they based, what are they known for?" and measure the response against ground truth.
This audit surface reveals four common failure modes:
- Conflation. The LLM confuses you with a similarly-named company. (We have seen this happen to a Singapore fintech confused with a US company 23 times out of 25 prompts.)
- Outdated information. The LLM describes a product you sunset in 2023 or cites a founder who left two years ago.
- Hedging. The LLM says "I don't have detailed information about…", which tells you your entity signal is weak enough that even the most advanced retrieval systems cannot confidently speak about you.
- Reframing. The LLM describes you accurately but with the wrong emphasis, positioning you as a commodity provider when you are a premium specialist.
Each failure mode has a specific fix on the entity-signal side. Wikidata updates, schema adjustments, canonical descriptions, directory reconciliation, and the audit prioritises them by impact.
Citation graph, where are LLMs actually getting their information about you?
Using Perplexity's visible citations and Workduo.ai's source attribution, we reverse-engineer the 30–80 external sources that LLMs are referencing when they discuss your category and your brand.
The output is a ranked list: which publications, directories, forums, review sites, and independent mentions are contributing most to how LLMs perceive you, and which are contributing to how they perceive your competitors.
Almost always, this reveals a tight cluster of 10–20 sources that punch disproportionately above their weight, and that cluster is the highest-leverage place to invest digital PR effort over the next quarter. Brands that spread PR thinly across 100 outlets almost never move the needle. Brands that earn four to six placements in the exact publications LLMs weight for their category move the needle in 60 days.
Technical accessibility, can the machines actually read your site?
The single most common root cause of poor AI visibility is, boringly, a misconfiguration that blocks AI crawlers from the site. We audit eleven technical layers:
- robots.txt directives for GPTBot, ClaudeBot, PerplexityBot, Google-Extended, Applebot-Extended, Bytespider, CCBot, Amazonbot, and OAI-SearchBot.
- WAF / bot-management rules (Cloudflare, Akamai, AWS WAF) that quietly return 403 to AI user agents.
- llms.txt presence, content quality, and version freshness.
- Server-rendered HTML completeness, does critical content survive without JavaScript execution?
- Structured data coverage. Organization, Service, FAQ, Article, Product, Review schema presence and validity.
- Canonical tags, sitemap completeness, and hreflang accuracy for multi-market brands.
- Response-code hygiene · 200/301/404 distribution, redirect chain depth, and orphan-page audit.
- Core Web Vitals, because slow-loading pages still correlate with lower crawl frequency.
- AI-bot rate-limiting configuration.
- Error-page content, whether 404/500 pages still return useful semantic markup.
- Canonical entity reconciliation across owned pages (About, Team, Locations).
A clean technical pass typically takes two to four engineer-days. The uplift it unlocks is often the largest single intervention in the entire audit, because without it, every subsequent investment compounds through a broken pipe.
Content architecture, is your content built to be extracted?
LLMs reward content that is easy to extract: clear question-answer structures, specific claims with citations, tables, lists, defined terms, and short paragraphs. We audit every cornerstone page against an extraction scorecard, heading hierarchy, FAQ schema coverage, claim density, source linking, paragraph length, and schema tagging.
Most brands discover two or three "hero" pages that are doing the heavy lifting, and dozens of supporting pages that are invisible to LLMs because they were built for human browsing, not machine extraction. Re-architecting three to five cornerstone pages to be maximally extractable typically produces the largest single-quarter GEO uplift we see.
Sentiment, what does the answer sound like when you are mentioned?
The final layer is qualitative. For every prompt where your brand appears, we score the sentiment and framing of the sentence the LLM constructs: positive, neutral, qualified, or negative.
Sentiment failures rarely originate in the LLM itself, they originate in the source material the LLM ingests. A single unresolved negative review ranked in Google's top 10 can reshape how four LLMs describe you for 18 months. A thread on a forum like Hardwarezone, a G2 review, or a Reddit post about your pricing can quietly anchor the tone of every AI answer about your brand.
The audit surfaces the specific source items that are shaping sentiment and prioritises them for response, resolution, or content counter-weighting.
What you walk away with.
A complete AI visibility audit at horatos.ai produces a 40–60 page deliverable containing:
- A single AI Visibility Score (0–100) benchmarked against your top three competitors.
- A prompt-by-prompt appearance map across all four major LLMs.
- A prioritised fix-list of 20–40 items, each tagged by impact (High / Medium / Low) and effort (Days / Weeks / Quarter).
- A citation-graph map naming the publications and sources to target for PR over the next two quarters.
- A technical accessibility scorecard with line-level recommendations for engineering.
- A content extractability report for your top 15 pages.
- A sentiment report with the specific third-party items shaping AI-answer tone.
- A 12-week roadmap sequencing the fixes for compounding impact.
This is the document that becomes the reference point for every GEO decision over the following year, whether executed in-house, with us, or with another partner.
When should you run an audit?
If you have not run one: now. The cost of the audit is negligible compared to the cost of running blind through another quarter of shifting AI-driven discovery.
If you ran one more than six months ago: now. LLM retrieval behaviour has changed materially in the last two quarters, new models, new grounding sources, and rapidly evolving citation logic. A six-month-old audit is already partially stale.
If you are about to launch a new product, enter a new market, or rebrand: before, not after. An audit reveals whether your existing entity signals will transfer cleanly, and in most cases surfaces two or three critical fixes that need to ship alongside the launch to avoid a six-month AI-visibility trough.
If you are an operator sceptical of GEO as a discipline: especially now. The audit is the objective, evidence-based artefact that either confirms or refutes the hypothesis. We have had brands run an audit expecting to disengage, and reallocate half their paid budget to GEO the same week. We have also had brands run one and rightfully conclude they had more urgent fundamentals. Either is a better decision than guessing.
