AI Visibility Audit: The 4% Problem

Search is turning into a shortlist

Buyers increasingly ask an AI assistant (ChatGPT, Claude, Gemini) who to hire instead of scanning the top-10 Google results.

The question is not whether you rank but whether you are on the shortlist it returns — and if not, who is.

Three assistants, one comparable test

The measurement is a Python script, run from the command line, that reads a prompt library, sends prompts in bulk to one model, and writes each answer back into that model's own column — three runs, one per model, produce the 120 answers.

Each provider exposes web search differently, so "with search on" means three different things.

# Anthropic — Messages API, server-side search tool
client.messages.stream(
    model="claude-sonnet-5", max_tokens=16000, system=SYSTEM_PROMPT,
    tools=[{"type": "web_search_20260209", "name": "web_search", "max_uses": 5}],
    messages=[{"role": "user", "content": prompt}])

# OpenAI — Responses API, `instructions` carries the system prompt
client.responses.stream(
    model="gpt-5.5", input=prompt, instructions=SYSTEM_PROMPT,
    tools=[{"type": "web_search"}])

# Google — google-genai, Search grounding
client.models.generate_content_stream(
    model="gemini-3.7-flash", contents=prompt,
    config=types.GenerateContentConfig(
        system_instruction=SYSTEM_PROMPT,
        tools=[types.Tool(google_search=types.GoogleSearch())]))

Findable by name, invisible among competitors

Three mechanisms behind the 4%

Who gets recommended instead

Not "competitors" in the abstract — the firms models actually name when asked for advice, and the specific claims they repeat about each.

The three assistants are not interchangeable

Visibility in one says very little about visibility in another — which is the practical argument for measuring more than one.

gpt-5.5 searched hardest — 4.3 distinct domains per answer against Gemini's 1.4 — and that alone explains all four unbranded mentions. The company's visibility currently rests on one model happening to dig deeper, which is luck rather than position.

gemini-3.7-flash is the alarming one. One citation in forty, zero unbranded mentions. It runs on live Google Search, so it is the closest available proxy for what Google's own stack knows — and by that proxy the company is close to absent from the ecosystem where most of its buyers decide. It is a proxy, not the AI Overview itself; that surface has no API.

The single clearest causal finding in the run came from the Ukrainian-language question. Both of its hits happened because a /uk/ version of the site exists; the model opened it, confirmed the page was genuinely in Ukrainian, ranked the company first, and — uniquely in all 120 answers — reproduced the phone number.

The fixes the data actually supports

Ordered so that nothing gets attributed to the wrong company before the rest is worth doing.

• Resolve the entity first. Sitewide LocalBusiness JSON-LD with exact NAP and sameAs; make the Wrocław address the primary one everywhere a machine reads it; put the phone in extractable markup. The Warsaw contamination and the 4% phone-recall rate are the same defect.
• Claim Oferteo and Fixly, then earn ratings there. The profiles matter less than what models quote from them — "5/5 from 45 reviews" is the currency of the shortlist, and 41 citations run through those two domains.
• Publish prices. Ranges per m² plus a worked 60 m² example. Prices appear in 8 answers and never from this site; in Ukrainian and Russian no source supplies a figure at all.
• Build estate landing pages. Olimpia Port, River Point, Ogrody Cesarski, Racławicka. The company already places sixth on the developer-state question purely on a passing mention of River Point; the firms above it name two or three estates each.
• Add a proof block to every key page. Founding year, warranty in months, insurance sum, projects completed, named project manager. Exactly the fields models already quote about competitors.
• Extend the Ukrainian section and add Russian. The one asset that demonstrably won answers, currently a translation of the homepage rather than a full section.

What this measurement cannot tell you

• It is one run. Web search introduces genuine run-to-run variance — different queries, different pages, different answers. These figures are magnitudes and stable patterns, not precise values. A second run of the same 120 cells would move the small numbers.
• APIs are not the consumer products. ChatGPT, Claude and Gemini ship hidden system prompts and their own retrieval stacks. Querying the APIs is the only way to hold conditions equal across vendors, and that comparability is bought by giving up literal fidelity to what a buyer sees in the app.
• Gemini with grounding is not Google's AI Overview. The block above Google's results is a separate surface with no API. Gemini indicates what Google's stack can find; measuring the Overview itself needs a search-results provider or manual capture.
• Mentions are counted, not ranked. Appearing sixth in a list counts the same as appearing first. At 4% that distinction is not yet the binding constraint, but it becomes one as the number rises.