Notes from a European Operator · AI Asset-Manager Observatory · No. 5 The share-of-voice readings behind these outliers are in the Index.
Brand size does not buy AI recommendation. The correlation between brand pull and AI share of voice is weak, and the outliers are the signal.
Brand size does not buy AI recommendation. That is the finding hidden inside the observatory’s newest cut, and it is the most useful number in the study for firms that do not have the biggest brand in the room.
I compared AI share of voice against a brand-pull proxy: English Wikipedia pageviews for the top 40 firms by AI reach, July 2026. The correlation in that cut is 0.25. Brand awareness explains almost none of who the machines recommend, and the panel is live if you want to read your own firm off it.
The proxy
Wikipedia pageviews stand in for web presence, with caveats
The comparison pairs the observatory’s AI reach (the share of answers naming each firm) with English Wikipedia monthly pageviews from the Wikimedia REST API, July 2026, for the 40 firms with the highest AI reach. It is a temporary proxy for firm-site web traffic, which is not wired in yet.
Wikipedia pageviews measure public interest in a firm, not traffic to its own site. The proxy is imperfect on purpose: it is the cleanest public signal available today, and the dashboard labels it as temporary. The shape it reveals is still the point.
A weak correlation is itself the finding. If brand awareness drove AI recommendations, the machines would hand buyers the same shortlist as the market’s share of voice. They do not, and the language cut shows how far the shortlist moves when only the language of the question changes.
The index re-reads this monthly, and the re-read holds the finding: the September 2026 reading puts the correlation at 0.35 across 33 firms, with JPMorgan AM at 13.7% AI reach on the second-largest pageview count in the set, and Robeco still overperforming at 36.1% on 908 views. The number moves with the panel; the shape does not.
The outliers
Brands without recommendation, recommendations without brands
| Firm | Wikipedia views | AI reach | Quadrant |
|---|---|---|---|
| JPMorgan Asset Management | 72,962 | 12.3% | Brand-led, GEO gap |
| State Street | 20,179 | 10.9% | Brand-led, GEO gap |
| PIMCO | 8,540 | 16.0% | Brand-led, GEO gap |
| Capital Group | 9,938 | 17.2% | Brand-led, GEO gap |
| Robeco | 908 | 42.3% | AI overperformer |
| Pictet | 2,900 | 43.9% | AI overperformer |
JPMorgan Asset Management carries the strongest brand pull in the dataset after BlackRock and Vanguard, and sits at 12.3% AI reach. Robeco, with 908 pageviews, reaches 42.3%. One firm is a household name that the machines barely recommend. The other is nearly invisible as a brand and recommended anyway.
The operator read
Brand awareness does not compound into AI share of voice
The machines recommend what they can find and cite, not what is famous. Grounded engines that search the web reward findable, quotable, structured content. Recall engines answering from memory reward training-data fame. Those are two different channels, and only one of them is content-shaped. The 10 that own 42% of AI mentions are in the short list.
That is the good news for firms outside the biggest names: the grounded channel is not budget-shaped. A smaller firm can out-recommend a giant with better findable content, which is exactly what the overperformers in the data are doing.
Fame still matters in recall mode, which is why the mode split is the strategic read: a recall advantage expires the moment the engine searches. And if the engines cannot read a firm’s site at all, no amount of brand awareness fixes the grounded question.
AI share of voice is a distribution metric, and the engine is a distribution channel. It belongs in the same distribution operating model as every other channel, with the same discipline: measure, find the gap, fix the source.
The operator question
Which quadrant is your firm in?
Plot your firm: brand pull against AI reach. If you are brand-led with a GEO gap, your content is the leak, and the fix is mechanical. If you are an overperformer, you are a primary source somewhere; defend that source and extend it.
The dataset is a proxy, but the shape of the market is not: the machines shortlist the findable, not the famous. What a single firm can do about that position is what the AI visibility audit measures, name by name.
Operator’s memo
Brands are remembered. Content is cited. AI share of voice is built from what the machine can find, not what it has heard of.
Availability
Not taking new engagements The studies and the library stay open, and questions about them get an answer.
Brand size explains part of this pattern and not the rest. Permission, answer-shaped pages and the language of the question carry the remainder, and the sequence I run across all four is ranked by cost with the numbers behind each step.
2 thoughts on “AI Share of Voice vs Brand Pull: The Outliers”