Do Not Hire an AI Visibility Vendor On Rank Alone

I coined Machine Relations because AI engines were already deciding who a market trusts, and almost nobody was measuring it. So when a vendor tells you an AI engine ranks them first for AI visibility work, I do not take the claim at the word. I check the grade behind it. Most of the time, right now, there is no grade to check.

The Machine Relations Index publishes a segment for exactly this: the category "AI Visibility & GEO" crossed with the question shape "how buyers choose." That is the segment an engine draws on when someone asks an AI assistant which AI visibility vendor or tool to pick. In the public view released 2026-09-28 (mri_score_v2.0+2026-09-28+054b584266b8, window 2026-05-10 to 2026-09-28), that segment cleared the evidence floor at 142 observed runs across 8 distinct dates, with 527 domain-run citations recorded inside it.

Here is what those 527 citations are actually made of.

The citation weight in a founder's hiring question

Confidence grade	Share of citation weight	What it means
Collecting (below evidence floor)	71.8%	Domain has not cleared 10 runs and 7 dates yet; rate can still move
C	16.4%	Thin evidence; directional only
B	5.7%	Stable enough to state as a fact
A	6.1%	Well-evidenced, high day coverage

Only 11.8 cents of every citation dollar in this question sits on evidence I would call load-bearing. The other 88.2% is a domain the model reached for once or twice, not a domain the record has watched long enough to trust.

Read the source-class split the same way. Of the 527 citations, 77.6% land on what MRI classifies as "other observed source": small tool sites, single blog posts, unclassified domains. Editorial publications hold 6.6%. Vendor-owned pages hold 5.4%. Analyst and consulting research, the category buyers usually assume is doing the vetting, holds one citation. Out of 527. That is 0.1%.

The two domains an engine reaches for most often when a buyer asks how to choose are YouTube (21.1% citation rate, confidence grade A, 30 of 142 runs) and a marketing-tools blog most founders have never heard of (13.4%, grade C). HubSpot and Semrush, the two recognizable vendor names that do show up, sit at 9.2% and 6.3% respectively, both grade B. No AI-visibility-specific vendor in this segment has reached grade A yet.

That pattern is consistent with how Google's own documentation describes AI Overviews sourcing: grounded on a range of web content rather than a fixed authority list, which is exactly why a citation set for a young question shape stays volatile until enough runs accumulate.

What this changes about the hiring decision

If a vendor pitches you on being "the top-cited AI visibility platform," ask which segment, which release, and which confidence grade. The honest answer today, for the exact question you are trying to answer when you hire, is that almost nine in ten citations backing any such claim have not cleared the floor that would make the claim durable.

That is not a reason to wait to hire. It is a reason to hire on something other than a ranking screenshot: a documented methodology, a named release you can re-pull yourself, and a willingness to show you the collecting-grade rows next to the graded ones rather than only the graded ones. The public MRI view publishes exactly that separation, segment by segment, because a number without its evidence floor is a number I would not build a company decision on.

I have watched founders make the SEO-era mistake of buying a rank instead of a capability. The AI-visibility hiring decision is repeating it faster, because the category is younger and the instruments measuring it are three months old, not fifteen years old. NIST's AI Risk Management Framework makes the same point about any measurement system: a number is only as good as the test set, method, and conditions behind it, and those have to be disclosed, not assumed. A grade of "collecting" is not a failure. It is the dataset telling you the truth before the market's story catches up to it.

FAQ

What does "how buyers choose" mean in the MRI dataset? It is one of six question shapes MRI tracks per category, alongside best tools, is it worth it, problem-first research, top lists, and comparisons. Each shape is measured separately because AI engines cite different domains depending on which question a buyer is actually asking.

Why is 72% of the evidence still "collecting"? A domain needs at least 10 observed citation runs across at least 7 distinct dates before MRI grades it A, B, or C. The AI Visibility & GEO how-buyers-choose segment itself just cleared the floor at the category level on 2026-09-28; most of the individual domains inside it have not.

Does this mean AI visibility vendor rankings are fake? No. It means most of them are early. A domain can be genuinely, frequently cited and still be graded "collecting" simply because the observation window has not run long enough. The grade is a statement about evidence volume, not about whether the citation happened.

What should I ask a vendor who cites a ranking? The release id, the segment (category and question shape), the confidence grade behind their specific number, and whether they will show you the same cut from machinerelations.ai/index so you can re-check it yourself.