AI Citations for Asset Managers: Who ChatGPT Cites, and Why

AI Visibility · Asset Management · ChatGPT Who the engine cites, in full, sits in the Observatory’s base run.

ChatGPT now cites about 21% fewer domains per answer, and its sources flow through three hidden pipelines, one of which carries nearly all instant-mode citations. For asset managers that is not a search story. It is a distribution story: the machine reads a short list of sources, and that list decides which managers get named.

Victor Roodhorst · roodhorst.eu · August 2026 · 12 min read

On March 4, 2026, OpenAI quietly switched the default model behind ChatGPT search. Nobody in financial services noticed. Within a week, the average ChatGPT answer cited 15 unique domains instead of 19. That is a 21% smaller citation pool, and it did not arrive as a slow drift. It stepped, on a Wednesday, when the model changed.

The finding comes from Resoneo and Meteoria, who monitored 400 prompts a day for 14 weeks and logged 27,000 comparable responses. Oncrawl's server logs corroborate it: ChatGPT-User crawls fewer pages now, and pages that still get visited are crawled less often. Unique URLs per answer fell from 24 to 19. The URLs-per-domain ratio never moved. ChatGPT did not stop reading deeply. It stopped reading widely.

For an asset manager, that number is not a curiosity. It is the size of the room your firm is allowed to stand in when a prospect asks a machine for a shortlist. And the room is not assigned by brand quality. It is assigned by infrastructure.

The mechanism

ChatGPT is not a search engine. It is a sourcing machine with three queues.

In July, Resoneo and Oncrawl published an analysis of 1,200 ChatGPT responses and 88,000 associated search results. They found three internal sourcing pipelines with working names: labrador, bright, and oxylabs. A separate line of evidence from Suganthan Mohanadasan, who read the ChatGPT web payload at packet level, found the same labels in a field called result_source, independently confirmed by two verifiers who read their own traffic.

Labrador is OpenAI's own web index, separate from Bing or Google. In instant mode it supplies nearly all citations. It carries full page titles and query-independent snippets of roughly 200 characters, usually built from the H1 and the text around it. Meta descriptions are ignored.

Bright and oxylabs are scraped web results, used when the query needs near-real-time data or labrador comes up short. Bright handles the bulk tier: commercial, shopping, finance, and weather queries. What the model sees there is thin: a URL, a title, a snippet of around 150 characters, and a date. Oxylabs skews regional and local.

Two more findings complete the picture. First, the mode decides the queue. GPT-5.3 Instant searches less and leans on its own index, which is the mechanical reason the pool shrank. GPT-5.4 Thinking reintroduces search fan-outs, but with site: operators aimed at trusted domains, so the pool may widen again while staying closed to newcomers. Second, specialized feeds run outside the web stack entirely: P1, P2 and P3 for shopping, and B1, B3, Yelp and Tripadvisor for local. If your business is retail banking or local advice, you are not in the web queue at all. You are in a local feed with different rules.

The source map

Which sources actually answer questions about European asset management

Resoneo and Oncrawl measured the machine. This section maps that mechanism onto the sources that actually feed AI answers about asset management. The mapping is our read of the observed pipelines, not Resoneo's data, so it is marked as interpretation.

Where European finance sources sit in ChatGPT's retrieval stack. Tiers are observed; the placement of each source type is this article's application of the mechanism.

Source type	Examples	Pipeline	What ChatGPT sees	What it means
Licensed global news	Reuters, WSJ, FT, AP, The Guardian, Axel Springer titles (Politico, Welt)	labrador	Full titles and excerpts up to roughly 1,080 characters	A mention here is a longer citation. Press is distribution infrastructure.
Reference and academia	Wikipedia, arXiv	labrador	Long excerpts; ~200-character snippets elsewhere	Your Wikipedia entity is a citation surface you do not control and must maintain.
Regulators and data providers	ESMA, ECB, FCA, EFAMA, Morningstar	bright (inference)	URL, title, snippet of ~150 characters, date	Compliance and data pages can be cited, but only the top of the page travels.
Corporate and brand sites	Your own domain and your competitors'	bright	Metadata only: URL, title, snippet, date. Your H1 is your citation.	The page competes as metadata. The answer must survive the H1 and first 100 words.
Community	Reddit, forums	bright	Metadata records; Reddit was the most-cited domain in one GPT-5.5 snapshot	For retail-facing funds, community threads can outrank brand content.
Local feeds	B1, B3, Yelp, Tripadvisor	outside the web stack	A different queue with different rules	Retail banking and local advice run here. Fund and institutional queries do not.

Read the "what ChatGPT sees" column twice, because it is the whole game. Under labrador, observed records carried up to roughly 1,080 characters of source text for news and reference organizations, which is consistent with OpenAI's publicly announced licensing deals with News Corp, Axel Springer, the Associated Press, The Guardian, and the Financial Times. Under bright, the record is a title, a snippet, and a date. One pipeline hands the model paragraphs. The other hands it a business card. That is not a credibility judgment. It is an infrastructure decision, made before any ranking or relevance step sees your content.

Now the uncomfortable part for European asset managers. The licensed tier is dominated by US and UK titles. The bulk tier is where your corporate site sits, and it is the tier that reads like a business card. And the community tier, which one snapshot found leading the cited-domain count on the newest model, is where retail buyers actually talk. Europe's managers are heavily represented in the middle tier, where the citation surface is thinnest, and thin where the excerpts are longest.

The grounded-versus-recall flip from my own observatory is the visible end of this: when the engine goes and reads the live web, it reads the tiers above, and BlackRock, a US firm, gets named first in 33% of answers to a question that explicitly asked for a European manager. Recall engines, working from an older, broader training diet, still lead with Amundi. The machine does not have a bias. It has a source list, and the source list is what you can change.

The ledger

The pool, measured before and after

The citation pool, before and after the March 2026 switch. Source: Resoneo and Meteoria, 27,000 responses.

Metric	Before	After
Unique domains per answer	19	15
Unique URLs per answer	24	19
URLs per domain	1.0	1.0

And the newest layer, supporting_websites, now shows which pages backed the same claim without winning the visible slot. Competition is claim by claim, not page by page. The wire shows the page that beat yours, and the pages you beat.

The diff

What changed when the model switched, and who is exposed

The data steps, it does not drift, and the steps follow model releases. Here is what the visible record shows.

The last row is a vendor study, 150 conversations on a single US Plus account, and it carries the conflict-of-interest flag that deserves: treat it as direction, not measurement. The first four rows are the corroborated set.

Now the part that connects to my own data, and the reason this piece exists. When I ran 487 answers across three engines in the English base run, the brand-pull read showed exactly which firms the shrinking pool will drop first.

AI reach vs brand pull for the eight names that matter most. AI reach: AI Asset-Manager Observatory, Q3 2026, share of 487 answers. Brand proxy: English Wikipedia monthly pageviews, July 2026. Directional, not a traffic census.

Firm	Region	AI reach	Brand pull proxy
Amundi	EU	72.5%	7,063
DWS	EU	56.9%	4,225
Schroders	EU	49.3%	5,313
Pictet	EU	43.9%	2,900
JPMorgan AM	US	12.3%	72,962
State Street	US	10.9%	20,179
T. Rowe Price	US	10.9%	7,364
Man Group	EU	11.1%	4,783

JPMorgan Asset Management carries 72,962 monthly pageviews of brand pull, ten times Amundi's, and reaches 12.3% of AI answers. Its citations were never as strong as its name. Those are the firms that lose first when the pool shrinks, because nothing about their underlying quality changed. The machine just stopped reading as widely.

The operator read

What an asset-management CMO does with this

Four moves, in order of leverage.

One: put the answer in the H1 and the first 100 words. Labrador builds snippets from the H1 and its surroundings and ignores meta descriptions. If your citable fact lives in paragraph twelve of a strategy page, the pipeline never carries it. The unit of AI citation is roughly a headline plus a short paragraph, so write your key pages as if that is all the machine will ever read. It probably is.

Two: know which queue you are in. Fund and institutional queries lean on labrador, where licensed news and Wikipedia carry the excerpts. Retail and ETF queries run through bright and the community tier, where your product page is metadata and a Reddit thread can outrank you. Retail banking and local advice run through the local feeds, which never touch the web stack. Each queue has different levers, and the difference is not optional detail, it is which content surface you are fighting on.

Three: treat earned coverage as citation infrastructure, not PR. The outlets with licensed long-excerpt pipelines, Reuters, the FT, WSJ, AP, Axel Springer titles, The Guardian, hand the model paragraphs while everyone else gets a business card. For a European manager, the FT and Reuters are the accessible front door to the longest citation surface that exists today. A mention there is not a vanity metric anymore. It is a longer citation, mechanically.

Four: re-baseline after every model switch, and measure as an interval. Your February citation numbers are wrong for the current model, and the next switch can move the pool 20% on a Wednesday. Run a fixed prompt set repeatedly, log every answer, and track citation share as a range over time, not a single check. That is exactly what the Observatory dashboard does every quarter, and it is why the movers panel exists: the first edition is the baseline, the deltas come after.

Fast answers

Questions I get about AI citations for asset managers

Is ChatGPT citing fewer sites?

Yes. After the March 2026 switch to GPT-5.3 Instant, average unique domains per answer fell from 19 to 15, about 21%, across 27,000 monitored responses from Resoneo and Meteoria. Unique URLs per answer fell from 24 to 19, and the URLs-per-domain ratio stayed flat.

Why do asset managers lose citations despite strong brands?

Because citations are assigned by infrastructure before relevance. Most instant-mode sources come from OpenAI's internal index, labrador, which carries roughly 200-character snippets built from the H1 and surrounding text and ignores meta descriptions. Commercial and finance queries run through a scraped tier, bright, that carries a URL, a title, a short snippet and a date. The Observatory data shows the pattern in action: JPMorgan Asset Management and State Street carry heavy brand traffic but single-digit AI reach.

Which sources does ChatGPT actually cite for finance questions?

The licensed news tier (Reuters, WSJ, FT, AP, The Guardian, Axel Springer titles) gets long excerpts through OpenAI's internal index. Regulators, data providers, and corporate sites sit in the bulk tier and are seen as a URL, a title, a short snippet and a date. Community sources like Reddit are rising: in one vendor snapshot on the newest model, Reddit was the most-cited domain. Local banking and advice queries run through separate local feeds.

Does Google ranking still matter?

Not the way you think. The overlap between Google rankings and ChatGPT citations is low, with median domain overlap around 10 to 15% in one study of 129,000 domains. What matters is presence in the sets the pipelines fetch: H1-level answers, licensed news coverage, and sources that repeatedly enter the retrieved set.

What is the single best fix for an asset manager?

Rewrite your key pages so the answer to your buyer's question lives in the H1 and the first 100 words, because that is roughly the unit labrador reads. Maintain your Wikipedia entity. Earn coverage in the licensed titles, because those get paragraphs while everyone else gets a business card. Then re-run the same prompt set every month and after every model switch, because the pool moves without notice.

Method, in brief

How the data was built, and what it does not prove

Every claim here traces to a published dataset or a named observation, and the confidence levels differ, so they are stated openly.

The 19-to-15 and 24-to-19 figures come from Resoneo and Meteoria: 400 prompts daily, 14 weeks, 27,000 comparable responses, published via Search Engine Journal, with Oncrawl server-log corroboration. The three-tier stack, the 200-character snippet unit, the ignored meta descriptions, and the shopping and local feeds come from Resoneo and Oncrawl's July analysis of 1,200 responses and 88,000 associated search results, reported via Abondance. The 88.1 / 9.9 / 1.7 / 0.3 share split, the long excerpts under labrador, and the supporting_websites layer come from Mohanadasan's single-account packet-level analysis, cross-checked by two independent verifiers; it is a snapshot of shape, not a population measurement. The first-party citation decline comes from a Writesonic vendor study of 150 conversations on one account, and is flagged as low confidence. The source map in section 02 applies the observed mechanism to European finance and is marked as interpretation: it is our read, not Resoneo's data. The asset-manager figures come from the roodhorst.eu AI Asset-Manager Observatory, Q3 2026: 487 answers, three engines, 20 frozen prompts, with English Wikipedia pageviews as a clearly labelled brand-pull proxy.

What this does not prove: labrador, bright and oxylabs are internal strings, and their names do not prove that Bright Data or Oxylabs the companies power them. No contract is visible in any traffic capture. The interior of the retrieval layer is a black box, the field values are undocumented, and they can be renamed or removed in any release. The mechanism is the finding. The vendor mapping is inference, and it is marked as such.

I re-run the Observatory every quarter and publish the movement. If you want the AI-citation reading for your specific segment, or want to compare notes on what actually shifts a grounded answer, I am at victor@roodhorst.eu. No form to fill in.

The firm-level version of that reading is the AI visibility audit: one competitor set, five languages, one scored benchmark.

Being cited is the durable half of this question. The volatile half is the shortlist position itself: when I replayed 600 saved answers, the first-place firm changed between runs in 48 of 60 prompt and engine cells. Citations compound. Positions flicker.

A shrinking pool changes the response. The first lever is the one nobody manages, and it is free: the four levers that move a firm into the retrieved set start with the crawler permissions that decide whether you can be fetched at all.

Victor Roodhorst publishes the AI Share of Voice Index for European Asset Management and runs roodhorst.eu, where he publishes original studies on how European financial services actually markets. He works from the regulated-operator seat, not the agency one.