Why does the AI visibility market have so many bad vendors?
Three conditions arrived together and each favours the seller. Buyers really have moved into AI assistants, and many brands really are invisible there. That fear is justified; the market selling the fix is young and its honesty has not kept pace.
- No standard metrics. Each vendor computes "AI share of voice" from its own prompt panel, so the numbers are neither comparable across vendors nor auditable by you.
- An invisible mechanism. Nobody outside the model labs can fully explain why an assistant named one brand and skipped another. When the buyer cannot verify the mechanism, sales copy fills the gap.
- Frightened budgets. Organic traffic is shifting, the board asks what ChatGPT says about the company, and "do nothing" feels riskier than "sign something".
Generative engine optimization is a real discipline. A measurable share of what is sold under the name is not. One disclosure: we sell AI visibility work ourselves, so read our bias accordingly. The article is written to be useful even if you never hire anyone.
Why can nobody guarantee an AI ranking?
Because the system itself rules it out, whatever the contract says.
Answers are sampled, not retrieved from a ranking. The same prompt, in two clean sessions on the same day, can name different brands in a different order. There is no index with slots, so no slot can be sold. Source work changes the probability distribution (how often you appear across many askings), never a fixed position.
Model versions reshuffle everything. Each release changes training data, retrieval behaviour and source weighting. G2's buyer research reported ChatGPT's share among B2B software buyers who use AI falling from 89% to 63% in a year while Claude rose from 1.4% to 18.5% (G2 via PR Newswire). "Ranking on AI" is several scoreboards, and all of them move.
What legitimate work does is raise the floor under that volatility: more independent sources a model can cite, consistent entity data it can ground against, presence on the surfaces it retrieves from. That is measurable and durable. It cannot fix a placement, and we never promise what an AI will say about you.
What five questions expose a weak vendor on the first call?
Ask all five. Each takes a minute.
- "Which prompt set, which models, how sampled?" A real operator hands you a written prompt list, names model versions and states runs per prompt. A weak one says "our proprietary tracking covers everything".
- "Show me baseline-to-delta on a past client." Pass: an anonymized before and after on the same frozen prompt set, variance included, with some prompts that did not move. Fail: a logo wall and "+340% AI visibility" with no denominator.
- "What is content, what is PR, what is entity work: what am I actually buying?" GEO bundles existing disciplines aimed at machine-readable surfaces. An honest vendor decomposes the bundle; a dishonest one says the algorithm does the work.
- "What happens to my results when the next model ships?" The only honest answer: answers will reshuffle, we re-baseline, and the durable layer is your sources and entity data.
- "Why can't I get this from a low-cost monitoring tool plus my content team?" Sometimes you can, and a serious vendor says so. One who cannot articulate value beyond measurement is selling a dashboard at retainer prices.
What are the 10 red flags, in vendors' own words?
1. The guarantee.
"We guarantee ChatGPT will recommend you within 90 days." Probabilistic systems do not issue guarantees; people who want your signature do. This sentence should end the call.
2. The proprietary algorithm.
"Our AI ranking algorithm has decoded how ChatGPT ranks brands." Nobody outside the labs has decoded model internals, and there is no stable ranking to decode. What vendors have is a prompt panel and a scraper: useful, but not secret physics.
3. The submission desk.
"We submit your brand directly to OpenAI, Google and Anthropic." No such desk exists. This is not an exaggeration; it is an invented mechanism.
4. llms.txt as a four-figure line item.
The file is plain markdown, takes about twenty minutes (how it works) and we have seen no engine document it as a ranking input. Publishing one is sensible; paying four figures for it is arbitrage on your unfamiliarity.
5. The share-of-voice chart with no methodology.
"You are at 4%; your competitor is at 31%." Ask which prompts, how many runs, which models, sampled when. If the pre-sale chart cannot answer, post-sale reports will not either.
6. Results inside one model cycle.
"You will see movement within 30 days." Source changes propagate through crawls, retrieval indexes and retraining over weeks to months. Anything that moves in days is retrieval noise or creative measurement.
7. No baseline before work starts.
A vendor who never captures a frozen baseline can never prove a delta, which is convenient for exactly one party.
8. The SOW that names nothing.
"Ongoing generative engine optimization, monthly fee." If the deliverable line has no nouns, no prompt set, no source list, no entity work, no re-measurement cadence, you are buying a subscription to vibes.
9. Everything, everywhere, one price.
"We optimize for all AIs." Each engine leans on a different source mix, and behaviour differs by language and market. Promising every engine with no prioritization means measuring none of them properly.
10. Content volume rebranded as GEO.
"Thirty AI-optimized articles per month." Engines reward citable authority, not throughput, and mass-produced AI text gives no other source a reason to cite you.
How do you judge a price without a market rate?
By the deliverables behind it, because published GEO pricing spans orders of magnitude for similar-sounding promises and no vendor survey is neutral. A fair quote can be read line by line:
| Engagement | What must be included | The version to refuse |
|---|---|---|
| One-time audit | Fixed prompt set, multi-model baseline, citation-source map, prioritized fix list | A templated PDF of scores with no prompt list |
| Entity and structured-data work | Named records created or corrected, with verification | "Schema optimization" with no list of what changes |
| Citation-source building | Which independent sources will exist at the end that did not exist at the start | Article volume with no external publication |
| Monthly retainer | Re-measurement against the frozen baseline, variance reported | "Ongoing optimization" with deliverables unnamed |
| Monitoring tooling | Fixed prompts, scheduled runs, multi-engine coverage | A resold tool inside a retainer, relabelled as proprietary |
For disclosure, our AI visibility audit starts from €490 and is a standalone deliverable, not credited toward packages. We publish it so you can compare scopes, not as a pitch.
What should a legitimate statement of work name?
A monthly engagement, if it is justified at all, reads like an engineering document. Five deliverables should appear by name:
- Prompt-set definition: the frozen list of buyer-relevant prompts (category, comparison, brand, adverse), agreed in writing before work starts.
- Baseline capture: multi-model, multi-run measurement of where you stand today, archived so neither party can move the goalposts.
- Entity and structured-data work: the specific records to be created or corrected, such as knowledge-graph entries, schema markup and consistent organization data. See Wikidata and knowledge graph work.
- Citation-source building: which independent, citable sources will exist at the end. This is the slowest line and the one that matters most.
- Re-measurement cadence: same prompts, same method, stated interval, variance reported, including the prompts that got worse.
A vendor who resists writing these down is telling you the deliverable is the invoice.
Where do Wikipedia and Wikidata fit into AI visibility?
Underneath the acronym GEO is mostly a sourcing problem: models cite what they trust, and they trust a short list of surfaces. Wikipedia is consistently among the most-cited, and Wikidata feeds the knowledge graphs that engines use to ground who you are. How the encyclopedia layer flows into answers is covered in Wikipedia AEO.
The test cuts both ways. A GEO vendor who never mentions sourcing, citation surfaces or entity work is selling content spam under a new acronym. And the Wikipedia services market has its own scam ecosystem, whose warning signs rhyme with these: guarantees, ghost agencies and unverifiable case studies. Same con, different surface.
When is a tool or an in-house team enough?
A retainer is the wrong purchase when your prompt universe is small (under roughly 100 buyer-relevant queries), you operate in one language, the engines already state correct facts about you and your content team can act on findings. A monitoring tool plus your own writers covers most of what a mid-range retainer claims; see AI brand monitoring tools.
External help earns its fee in three situations: multi-market measurement in languages you do not staff, entity and citation-surface work your team cannot do in-house (knowledge-graph corrections, encyclopedic sourcing), and active errors about your company that must be fixed at the source. All three are projects with ends, which is why fixed scope usually beats an open-ended retainer.
Frequently asked questions
No. Guarantees about a specific AI answer contradict how language models work. Legitimate vendors promise a method, a measurement and an agreed prompt set, and report probabilities with variance.
Source changes usually need weeks to months to reach crawls, retrieval indexes and later model versions. A claim of visible movement in days is more likely sampling noise than effect.
One that names a frozen prompt set, a baseline, entity and structured-data work, citation-source building and a re-measurement cadence. If a line has no noun, ask for one.
Publishing one is cheap and sensible, but it is a small file, not a service line. Treat a four-figure charge for it as a sign that the rest of the proposal needs scrutiny.
Yes, partly. Ask for the prompt list and run counts, then rerun a sample of the prompts several times in clean sessions. If your results differ widely, the number is noise.
Not always, but Wikipedia and Wikidata are among the surfaces models lean on. A notable company with weak encyclopedia presence often has a sourcing gap; a notability audit shows whether a page is realistic.
Fixed scope fits audits, entity work and one-off corrections. A retainer only makes sense when the SOW names recurring deliverables and measurement against a frozen baseline.
How WikiBusines helps
We do the source and entity layer, with scope named before you pay:
- AI visibility audit, from €490: multi-model baseline and citation-source map; a standalone deliverable, not credited toward packages.
- Wikidata entity, from €680: structured entity records that engines use to ground your brand.
- Notability audit, from €490, and English page creation, from €2,400, 4-5 weeks: for brands whose gap is the encyclopedia layer, with paid editing disclosed in WikiBusines' name.
To scope an audit against a proposal you already hold, contact us.
Authors and editorial policy
Bohdan DubylovskyiCo-founder and CBO, WikiBusines→
Roman MelnykCo-founder & Chief Editor, WikiBusines→The article was prepared by the WikiBusines co-founders. We sell paid Wikipedia services, so we disclose this interest: paid editing is always declared on Wikipedia in WikiBusines' name, and claims about the rules are backed by links to primary sources.
Editorial policy: payment disclosure, conflicts of interest, sources