Disclosure up front. We are Proofmend. We sell generative engine optimization work, which makes us one of the names this question is about. That is exactly why the piece below hands you a ruler rather than a shortlist. Use it on us too.
Every week we put buyer-shaped questions to nine engines and store the answer next to the conditions it was produced under. The provider-recommendation questions are the least stable ones we track. The same prompt, a week apart, returns a different set of names, and several models now annotate vendor listicles as vendor material rather than treating them as evidence. A ranked list of firms is the one artifact this category cannot reliably produce, and it is the artifact everyone publishes.
Where your shortlist actually comes from
Five channels supply nearly every candidate list, and each one selects for something that is not competence.
| Where the name came from | What that channel selects for | The failure it tends to produce | Minimum check before you book a call |
|---|---|---|---|
| A "top 10 agencies" article | Whoever paid for or wrote the placement | The list and the listed firms are the same people | Look for the disclosure line, then check whether the author sells the category |
| An assistant's own answer | Whichever domains were retrievable that day | Names drift week to week, and small firms with no third-party record never appear | Ask the same question twice, three days apart, and compare |
| A referral from a peer | Fit with someone else's stack and someone else's problem | A team excellent at content work handed a technical entity problem | Ask what specifically was delivered, not whether they were happy |
| A directory or review platform | Willingness to maintain a profile, and review volume | Profile age beats current capability | Read the negative entries only, they carry the detail |
| Inbound outreach quoting your ranking | List-building software | The audit attached was generated in bulk before anyone looked at you | Check whether the audit names a page on your site that actually exists |
None of those channels is worthless. Each one is a sampling method with a bias you can correct for once you know what it is.
Why nobody can be best at the part you care about
The outcome you want lives on the platform side. Whether an assistant names you tomorrow depends on retrieval, on which sources it pulled, and on model behaviour that changes without notice. Our own weekly records show the same question producing different citation pools within days, with no intervention from anyone.
So the honest claim any provider can make is bounded to what they control. Written deliverables, a defined scope, checkable acceptance points. Proofmend publishes that boundary on its own pricing page, in the same words for every market: no guarantee of rankings, AI mentions or citations, traffic, leads, or revenue. A firm that will not put a comparable sentence in writing is selling you the part of the job that is not theirs to sell.
Five artifacts, requested on the first call
Ask for these five things before scope, before price, before case studies. The response tells you more than a portfolio does.
- Raw records. Question text, engine, mode, timestamp, the full answer, and the source list it quoted. If only conclusions come back, nobody can reproduce the finding later, including the next agency you hire.
- A written unknown policy. Ours logs anything untested or unavailable as unknown, with no estimate. A report containing zero unknowns has usually written estimates in as facts.
- The delivery clock definition. Does the period start at signature or when materials, access and review are actually available? This single clause decides who owns a slipped schedule.
- Deliverables in countable units. Work orders and acceptance points can be checked. Person-days and article counts measure input, which is a legitimate purchase, but a different one.
- A price you can see. Ours is public: US$699 for a seven-working-day Evidence Audit with no implementation included, US$1,380 for a thirty-day engagement covering at most six agreed work orders, and US$6,980 for a ninety-day programme, half due at start and the rest split across two acceptance gates. Those figures carry conditions, a limited test price for the first three paid design partners and an exclusion list covering medical, financial, admissions and franchise-recruitment work for the first ten customers.
Score each provider zero or one on those five. In our experience the spread separates candidates faster than any capability claim, because four of the five are answerable in a sentence by anyone who runs the work honestly, and unanswerable by anyone who does not.
One number to keep in your head while you shop
In our deduplicated dataset of 187,818 citations, domains owned by the brand itself account for 1.37% of everything quoted. Your own site governs whether the description of you is accurate. Almost everything that decides whether you are named at all sits on pages you do not own.
Any provider whose entire plan happens inside your CMS has quietly narrowed the job to 1.37% of it. Any provider who claims control over the other 98% is describing something they cannot deliver. The workable engagements sit between those two positions, and asking where a firm sits on that line is a faster filter than reading another ranked list.
Written by the Proofmend team. Prices and the outcome boundary read from our own pages on August 26, 2026, at https://proofmend.com/pricing/ . For which crawlers fetch pages on behalf of assistant answers, the operators publish their own documentation, for example https://platform.openai.com/docs/bots and https://docs.perplexity.ai/guides/bots .
Top comments (0)