AI Doesn't Have a Rank. It Has a Consideration Set.
Rand Fishkin's SparkToro research just proved AI brand recommendations are random more than 99% of the time. The industry called it a tracking problem. They missed the actual diagnosis.
Rand Fishkin did the study nobody had done. He went in as a skeptic — convinced AI brand tracking was a waste of budget — and the data confirmed his skepticism harder than expected.
The SparkToro research, co-authored with Patrick O'Donnell of Gumshoe.ai and published in January 2026, ran 2,961 prompt tests across ChatGPT, Claude, and Google AI with 600 volunteers across 12 product and service categories. The question: are AI recommendation lists consistent enough to track?
The answer was no. Not marginally no. Structurally no.
Ask ChatGPT to recommend brands in a category 100 times and you have less than a 1-in-100 chance of seeing the same list twice. Want the same list in the same order? That's roughly 1 in 1,000. The variables — which brands appear, how many appear, in what sequence — change on every single run.
The industry's response was: this is a measurement problem. Upgrade your methodology. Run 60–100 prompts per query. Track visibility percentage, not rank. Trend the data over time.
Technically correct. Also the wrong diagnosis entirely.
AI Recommendation Works in Two Layers — and the Industry Is Only Watching One
Here is what the variability actually reveals.
AI systems are probability engines, not databases. Each response is a weighted draw from what the model has learned to associate with a topic. The list changes because the selection is probabilistic — the AI draws from a pool of entities it considers relevant, not from a ranked queue.
Two distinct things are happening, on completely different timescales:
The consideration set is relatively stable. Across the SparkToro research, the top brands in each category appeared in 55–77% of responses regardless of how prompts were phrased. Sony, Bose, and Apple showed up across nearly every headphone recommendation run. Ramp showed up consistently in B2B fintech. The AI had enough corroborating signal to include these brands reliably. The pool is probabilistic at the edges, but the core is stable.
The rank within that pool is effectively random. The AI doesn't maintain a hierarchy. It draws, every time, from the consideration set. As Fishkin concluded: "any tool that gives a 'ranking position in AI' is full of baloney."
This distinction changes everything about where resources should go. The industry built a $100M+ tracking infrastructure — Search Engine Land estimated the spend — premised on the idea that rank matters. The research is saying rank doesn't exist in the way anyone thought it did.
What Controls Whether You're in the Consideration Set
The brands at 70% inclusion in the SparkToro research do not have better SEO or tighter schema markup than the brands at 10% inclusion. They have deeper presence in the publications, forums, and indexed sources that AI engines use to form confident judgments.
The data on this is consistent across every independent study that has looked at it.
Ahrefs' analysis of ChatGPT's most-cited pages found 65.3% come from domains with a Domain Rating above 80. High-authority sources. The AI isn't pulling from your website's blog — it's pulling from publications it has been trained on for years.
Muck Rack's AI reading study — analyzing over 1 million AI prompts — found 85% of non-paid AI citations come from earned media sources. Not owned content. Not press releases. Editorial placements in journalism outlets.
The Fullintel-UConn study presented at the Institute for Public Relations Research Conference found 47% of all AI citations in direct responses came from journalistic sources, with 89% of links cited being earned media and 95% being unpaid coverage.
A Moz analysis of 40,000 queries found 88% of AI Mode citations do not appear in the organic top 10. The AI is drawing from a different universe of sources than traditional search ranking.
SE Ranking's study of 10,000 keywords in Google AI Mode found only 9.2% URL consistency when running the same query three times on the same day. In 21.2% of cases, zero URL overlap between response sets. The AI is pulling from a broad pool, not a fixed list — and that pool is determined by domain authority and editorial credibility, not optimization score.
The picture is consistent: consideration-set inclusion is driven by authoritative, independent, earned coverage in high-DA publications. Not by anything that happens inside your own properties.
Why "Better Tracking" Won't Fix the Underlying Problem
The tracking tool market will upgrade its methodology in response to the SparkToro findings. More prompts per query. Visibility percentage instead of rank position. Trend lines instead of snapshots.
That is the right response to the measurement problem.
But if the strategic implication is "measure your visibility percentage and optimize accordingly," the strategy that follows is still wrong — because visibility percentage is downstream of consideration-set inclusion, and consideration-set inclusion is downstream of earned authority, not prompt engineering.
Here is the cleaner way to think about the two strategies:
| If you're solving the measurement problem | If you're solving the presence problem |
|---|---|
| Run 60–100 prompts per query | Earn placements in publications AI engines trust |
| Track visibility % over time | Track your editorial footprint across Tier 1 media |
| Test prompt variations | Build entity signals across independent domains |
| Upgrade your tracking tool | Invest where AI citation actually originates |
| Know how often you appear | Control why you appear |
The Princeton/Georgia Tech GEO research (Aggarwal et al., SIGKDD 2024) found that adding statistics to content improves AI citation probability by 30–40%. That matters. But the citations the AI pulls with higher frequency are still coming from the high-authority publications first. Structure amplifies what authority already established.
Amanda Natividad's follow-up analysis at SparkToro found that across 142 human-written prompts for the same underlying intent, the average semantic similarity score was 0.081 — meaning prompts were as different from each other as "Kung Pao Chicken and Peanut Butter." Yet despite the prompt chaos, the underlying consideration sets held. Brands in the set showed up regardless of how the question was asked. Presence is the variable, not prompt format.
AuthorityTech's own research tracking citation rates across earned vs. owned distribution found that earned media placements generate 325% more AI citations than equivalent owned content. Same brand, same information, different origin. The AI reads the editorial record.
The Actual Implication for Founders
You cannot optimize your way into the consideration set from inside your own properties.
Structured data, schema markup, and on-page optimization matter — they're table stakes for basic crawlability. But they don't build the AI's confidence in your brand. That confidence comes from external, independent, authoritative sources that the AI has seen corroborate each other over time.
Brands that appeared in 97% of AI responses in Fishkin's research had one thing in common: dense, consistent coverage from multiple independent directions. The AI had seen them discussed authoritatively from many angles. It had confidence. Brands appearing in 5–10% weren't penalized — they were simply not validated enough for the AI to commit.
This is why Bain's 2025 AI search study — finding 80% of search users now rely on AI summaries at least 40% of the time — is actually a measurement of a structural shift in how brands get considered at all. The AI is making the first cut. If you're not in its consideration set, you don't get a first-page problem. You get a non-existence problem.
Machine Relations is the name for this entire system: the discipline of building your brand's presence in the sources AI engines already trust, so that when a prospect asks any AI system who leads your category, your brand is in the frame the AI draws from.
PR's original mechanism — earning placements in credible publications through real editorial relationships — turns out to be exactly the mechanism AI engines use to decide who's worth citing. Same mechanism, new reader.
The SparkToro research proved AI rank is noise. What it pointed at, without naming directly, is that the thing that drives consistent AI citation is the same thing that drove consistent human brand recognition before AI showed up: third-party, authoritative, independent coverage in publications the reader has been trained to trust.
That's not a tracking problem. It's a presence problem. They have different solutions.
See where your brand currently lands in AI consideration sets at app.authoritytech.io/visibility-audit.
Related Reading
- Machine Relations by Industry: AI Visibility Playbooks for 2026
- AI Visibility for Media & Entertainment Companies: The 2026 Earned Media Playbook
Jaxon Parrott is the founder of AuthorityTech and the originator of the Machine Relations discipline.