Can you buy AI visibility on price, or is there real differentiation between vendors?

Real differentiation exists, and buying on price alone will not find it, because the cheaper quote is usually buying you fewer measurements rather than the same measurements for less money.

Two things make a straight price comparison hard in this category. The first is that nobody has agreed what the unit is. One vendor sells prompts tracked per month, another sells engines polled per week, another sells a seat in a dashboard, and a fourth sells a programme of page and off-site work with monitoring attached. A monthly figure next to another monthly figure tells you almost nothing until you have converted both into the same unit.

The second is that the headline number each vendor reports is produced by its own method, and the methods do not agree. According to Deepak Gupta's buyer's guide for 2026, which evaluated eighteen vendors that claim to track brand visibility inside AI engines, one reported 12% share of answer where another reported 31%, for the same brand on the same date, with both figures defensible given different engine coverage, prompt sets, sampling cadence, citation parsing and persona settings. His conclusion is that triangulating across vendors is mathematically meaningless, which is a strong claim and a correct one.

So the honest shape of the decision is this. Differentiation is real but it sits in the measurement design and in how much work gets done on and off your site, neither of which appears on a rate card. Price is a fine tie-breaker once two offers do the same thing. It is a bad first filter, because the thing it selects for is a smaller sample.

Two offers that look five hundred pounds apart are often a long way further apart than that, in opposite directions, because one is sampling four engines weekly and executing page changes while the other is sampling one engine monthly and emailing you a list.

Why can you not compare two AI visibility vendors' numbers?

Two vendors' numbers are not comparable because each one is an estimate from a sample the vendor designed, and the samples differ in ways that move the result more than your website does.

Start with how unstable the underlying thing is. Ahrefs tracked about 37,500 citations across Google AI Mode, ChatGPT, Google AI Overviews, Gemini and Copilot, asking the same questions daily for twenty days, and found that only 6.6% of day-one AI Overviews citations and 4.0% of day-one ChatGPT citations survived all twenty days, while between 50% and 83% of day-one citations were gone by day two depending on the tool (published 29 October 2026). A vendor that samples once a week and a vendor that samples daily are not measuring the same population, and the weekly one will report a calmer picture for reasons that have nothing to do with your pages.

Then add coverage. A vendor polling four engines and a vendor polling eight will produce different shares for the same brand, and even inside Google the surfaces disagree. Ahrefs compared 540,000 matched query pairs from September 2025 in the United States and found only 13.7% of citations overlapping between AI Overviews and AI Mode, rising to 16.3% across the top three citations, while the answers themselves matched at 86% average cosine similarity (published 15 December 2025). Counting one of those two surfaces and calling the result your Google visibility is a methodological choice that a rate card will not show you.

The practical consequence is narrow and useful. Ask each vendor for its engine list, its polling cadence, its prompt set and whether it re-baselines when it changes the method. Then pick the methodology you are willing to live inside for a year, because the one thing you cannot do is average two vendors' answers.

What are you actually paying for in an AI visibility engagement?

An AI visibility engagement is paying for three separable things, and vendors differ most in how many of the three they actually do rather than in the quality of any one of them.

The first is measurement: a prompt set, a cadence, a record of which engines cited which URLs, and a read on whether a change is real or inside the noise. The second is page work: rewriting titles, headings and opening sentences so they match the questions buyers ask, fixing what blocks a crawler, and making the answer extractable. The third is off-site work: getting named on the third-party pages and in the places that shape what a model has learned about your category.

The weight of evidence says the third is where recommendation is decided and the second is where retrievability is decided. Mahe Chen and Nick Koudas analysed 158,099 generative-engine answers carrying 468,037 citations across two 90-day collections and found that which pages the engine's own search retrieves explains about 32% of the variation in citation while page properties explain 2.8%, and that ChatGPT names brands with no supporting page at all because, in their summary, pages are cited when searched and brands are also named when remembered (arXiv 2610.15033, submitted 10 October 2026).

Read that as a buying instruction rather than as a reason to despair about page work. A vendor selling only a dashboard is selling you the first of the three. A vendor selling page rewrites is selling the second, which is the part that makes you eligible. A vendor that has no answer at all for the third is quoting for a smaller job than the one you have, and its lower price is accurate.

Our own published results page is deliberately explicit about which of these a score movement proves: an audit score that moved from 17 to 68 on an early client is an input metric, and share of voice is the outcome metric that takes months and is tracked separately.

Which differences between vendors are real, and which are packaging?

The real differences between vendors are the ones that change what gets measured or what gets changed on your site, and the packaging differences are the ones that change how the same work is described.

Gupta's guide proposes four dimensions to score every shortlisted vendor on: engine coverage including region and persona handling, methodology transparency including whether changes are re-baselined rather than silently applied to history, data exportability including raw answers and timestamps, and pricing transparency including what happens at renewal if usage doubles. Those four are a good filter precisely because a vendor cannot answer them with positioning.

What varies	Real or packaging	How to test it in one question
Engines polled, and how often	Real	Which engines, at what cadence, and what does one run cost you?
Prompt set design	Real	Who writes the prompts, and can we add and remove them?
Whether anyone edits your pages	Real	Who makes the change after the report, you or us?
Off-site and third-party work	Real	What is in scope beyond our own domain?
Raw data export	Real	Can we export the cited URLs, prompts and timestamps?
The name of the score	Packaging	What is the formula, and is it published and versioned?
Dashboard design and alerting	Packaging	What decision does this screen change?
The label AEO, GEO, AI SEO or LLM SEO	Packaging	Nothing. Ask about the four rows above instead.

Score names belong in the packaging column for a measured reason. Elisha Bajemon and Andre-Louis Rochet gave ten answer engines the same query and the same five candidate sources, so that no engine retrieved anything of its own, and found the engines agreeing with each other at a Spearman correlation of 0.704 while the best available scorer reached 0.575 to 0.578 (arXiv 2610.38090, submitted 28 October 2026). The authors disclose that they are employed by TW3 Partners, which sells generative engine visibility services, and they state plainly that their setup does not establish that the ranking transfers to engines that search the web. Both halves matter: scoring a page is meaningful once a page is in the candidate set, and no vendor's score, including ours, should be sold as a prediction of what engines will do.

How do you run a price comparison that is not meaningless?

A price comparison becomes meaningful once both quotes are reduced to the same unit, and the unit that works in this category is cost per tracked prompt per engine per period.

Do the conversion before you look at the monthly totals. Ask each vendor for the number of prompts, the number of engines, the polling cadence and the contract length, then divide. Gupta's guide recommends exactly this reduction and adds a second question that is easy to forget, which is what the renewal looks like if your usage doubles. A quote that is cheap at fifty prompts and punitive at two hundred is a different offer from the one you compared.

Then separate the measurement line from the work line, because they scale differently and a blended monthly figure hides which one you are buying. Measurement cost rises with prompts, engines and cadence. Work cost rises with the number of pages touched, the number of approval cycles your organisation requires and whether the vendor is executing or advising. Two quotes with the same total can be ninety per cent monitoring and ninety per cent execution respectively.

Finally, price the thing you will do with the output. A report nobody acts on costs more than its invoice, because the work it recommends never happens and you pay again next quarter for the same finding. A vendor whose deliverable is a list of recommendations your team has no capacity to implement is selling you a cost rather than an outcome. Ask who makes the change, and ask it before you ask the price.

If you want a reference point that costs nothing, a free visibility assessment runs your buying questions through ChatGPT, Perplexity, Google AI Overviews and Claude and shows you which pages the engines answer with today, which is the same baseline any paid vendor would start from.

When should you just buy the cheapest option?

Buying the cheapest option is the right call in three situations, and recognising them will save you from paying for a programme when a subscription would have done.

Buy cheap when you are still establishing whether anything is happening at all. If you have never looked, a low-cost monitoring tool on a narrow prompt set answers the only question you currently have, which is whether buyers in your category get an answer that names anyone. Our own category research exists because that question is worth answering before anything else: across eight pharmaceutical skid categories, no fabricator was named by all four engines in any contestable category we measured, which is the kind of finding that changes what you do next.

Buy cheap when the scope is genuinely one site and one product line, your pages are already yours to edit, and you have somebody who will do the editing. The page-level work is learnable, and the published guidance from Google is less demanding than the category implies: its own documentation on AI features says a page needs to be indexed and eligible for a snippet, and that there are no additional technical requirements, no new machine readable or AI text files, and no special schema.org data to add.

Buy cheap when the decision is reversible and small. A twelve-month commitment on a method you have not tested is the expensive mistake in this category, not a monthly tool.

Do not buy cheap when the gap is off your own site, when several teams have to approve every page change, or when the answer has to hold up in front of a board. Those three make the sampling design and the execution capacity the whole value of the engagement, and they are the things a low price removes first. SIGNALS is one of the firms that sells the programme version of this work, which is a reason to check our account of ourselves against the vendors who sell the tool version.

What else do people ask about comparing AI visibility vendors?

People ask these five questions about comparing AI visibility vendors while trying to work out whether two quotes in front of them describe the same job.