Sizing the taste market

The specific market — human aesthetic preference and creative evaluation sold as a service in 2026 — cannot be sized from public information. No vendor breaks out creative revenue. No lab discloses spend. No analyst series covers the vertical. Anyone quoting a figure is extrapolating, including this page.

What can be done is to bound the containing market, apply a defensible share, and cross-check from three other directions. Here is the working, with every input visible so a reader can disagree with a specific step rather than with the conclusion.

#Step one: the containing market has a verified floor

Mercor names its customers — OpenAI, Anthropic, Google DeepMind, Reflection AI and Thinking Machines Lab, with Meta suspended after a March security incident — which is what makes the 91%-from-labs figure usable. Expert data for frontier labs carries the full picture of that buyer set.

#Step two: the creative share is 1–3%, and the justification is an absence

Mercor's disclosed domains are "lawyers, doctors, writers, and researchers." No design or visual-arts vertical is named at any major vendor. Micro1, Surge and Handshake coverage foregrounds law, medicine, finance and software. Handshake's top rates — $300–500/hr — go to physicians, STEM researchers and radiologists, and it has no design vertical at all.

Creative work is plainly being done — Mercor runs a creative vertical with live design postings, as The generalists in creative documents — but it is not large enough for any vendor to name. That supports a share of 1–3%, which against a $6–10B market gives $60M–$300M.

#Step three: three cross-checks

Bottom-up from the labs. The realistic buyers of aesthetic preference data number roughly 17: OpenAI, Google DeepMind, Black Forest Labs, ByteDance, Kuaishou, Alibaba, MiniMax, Midjourney, Runway, Luma, Ideogram, Recraft, Krea, Stability, Suno, ElevenLabs, Udio. Imagen 3's disclosed campaign was 366,569 ratings from 3,225 raters (arXiv 2408.07009). At crowd rates of $12–16/hour and roughly 100 ratings an hour, that campaign costs about $44K–$59K. Twenty such campaigns a year across 17 organisations is $15–20M a year at crowd prices. At professional rates of $50–250/hour the same volume is $60M–$400M — but only under total conversion to professional panels, which nothing in the record suggests is happening.

Bottom-up from the app layer. Canva is the only app-layer company with a confirmed "human data spend" line, and its supply is largely in-house teams in the Philippines and the EU (The creative tool layer). If Canva spends single-digit millions and is the most advanced buyer in its cohort, the whole layer plausibly spends $10–50M/year, mostly on salaries. The externally addressable portion today is likely under $10M.

From the vendors selling it. The three companies explicitly selling creative taste have raised $7.9M (Intelligence), $18.5M (Taste Labs) and an undisclosed amount inside a $45.2M parent (Contra Labs). Seed-stage vendors with no disclosed customers do not indicate a large realised market. AfterQuery's $100M+ is generalist expert data, and its design work was a free-panel marketing exercise.

#The answer

For scale: that whole market is smaller than Arena's single-company annualised revenue of $100M, and roughly one twentieth of Mercor's gross run rate. A vendor taking 20% of the central estimate is a $20–30M revenue business — a real company, and not the one the "Mercor for design" framing implies. How much money is actually in the buyer pool and The taste read carry what follows from that.

#The pricing contradiction at the centre of the business case

Two structural notes on why the gap has held. First, the professional input labs do buy sits at the specification layer — GenAI-Bench's prompts came from professional designers and were then judged by crowds, a split that recurs everywhere (The generative-media labs). Second, the highest-credential panel in the whole literature was also the cheapest, because UI-Bench's 194 experts were invited rather than hired. Expertise has been available for free to anyone with a plausible paper, which is a supply fact the pricing model has to survive.

#The reference price for a field-defining benchmark is four figures

HEIM evaluated dozens of text-to-image models across alignment, quality, aesthetics and originality. It cost thirteen thousand dollars. GenAI-Bench cost roughly $9,600 — about 800 annotator-hours at $12/hour — and Google DeepMind then used its designer-sourced prompts to evaluate Imagen 3. The single most consequential prompt set in commercial image evaluation cost less than a mid-market SaaS seat contract.

The same anchor from the professional side: the Contra × Lica TASTE study paid 10 designers a flat fee averaging ~$90/hour for 13–16 hours each — about $13,000 for 14,400 comparisons, or ~$0.91 each (arXiv 2605.20731). Extrapolated, a million-comparison frontier-grade corpus is ~$1M of direct annotator labour. That is a fundable line item for any lab with a program manager, not a moat — which is why the sellable unit has to be the thing that does not scale linearly: rubric design, calibration, named panels, narrated process. Eight units and one cost anchor and The trajectory moat carry that argument.