Mindshare and Mentions Are Not Substitutes for Trader Positioning

If you are shortlisting a social data vendor, settle first which of four different products you are buying. They get compared as though one replaces another, and they are complements. A team that needs one and buys another ships the wrong feature and finds out a quarter later.

The scale of the gap: in the seven days to 9 September 2026, the accounts and channels Centaur monitors published 125,928 messages. 988 became a new position, and 711 of those came from a post where the trader stated the position they were opening. A product built to measure the 125,928 and a product built to extract the 711 do not compete for the same slot in a stack.

Four questions, four different units

Separate these products by what one row of their output actually is. The row shape determines every join, every query and every feature you can build on it.

Those definitions are the vendors' own and they are sharper than the category names. Kaito describes mindshare as a metric quantifying "the prominence of a specific project within the overall crypto market", weighted by account reputation rather than by raw impressions. Messari sells the same measure under Signals, defining it so that an asset with a 1% mindshare accounts for 1% of conversations. Santiment defines social volume as a count of documents containing a search term, so a message naming an asset ten times increments it by one.

Those are different objects. Mindshare and social volume are aggregates over a body of text, so the row is a bucket and a source can only ever attach to the bucket. A mention is a document. A position is an event with a start, a side and an end. You cannot compute the fourth from the first three.

One disclosure before the argument. This is published by Centaur, which occupies the fourth row, so read the placement with the scepticism it deserves. Every description above is the vendor's own words, and the claim is that the four are complements rather than that one of them wins. The definitions themselves sit in what social trading data is. This piece is about which of them to buy.

Direction is where they separate

Direction is the field most buyers assume they are getting, and it arrives in three forms.

Magnitude without a sign. Mindshare and social volume tell you a name is loud, not which way the people making the noise are leaning, because a mention has no side. Elfa AI is the most direct of any vendor about that boundary, telling developers its measurement endpoints "measure attention, not meaning".

A sign without a lifecycle. Stance products do carry a side. Stocktwits users tag each post bullish or bearish themselves, and the tag aggregates into a per-symbol score every five minutes. Stocktwits is careful about what that is for, saying in its own help centre that sentiment "isn't designed to predict future price movements". A stance attaches to a post. It has no entry price and no end, so there is nothing to mark against the market later.

A sign attached to an instrument, a price and a lifecycle. The twelve most-opened assets across the same seven days, with the side on every row:

Two things there are invisible to an attention product. Memory and semiconductors carry no disclosed short side while Pons splits almost evenly, and those are different market conditions that can produce identical mention counts. And seven of the twelve rows are equities, ETFs or an index, because the accounts posting crypto trades are frequently the same accounts posting stock trades, so a crypto-only attention feed measures a subset of its own population.

MicroStrategy makes the point on one name. It appears in crypto discussion constantly, since its investment case is a bitcoin proxy. In the 90 days to 28 October 2026, tracked traders opened 14 MicroStrategy positions, 10 long and 4 short. A volume of mentions has no sign, and a stance tag has one with nothing to settle it against.

What one row looks like, and how long it lives

At 16:37 UTC on 3 September 2026, Ariel Hernandez, posting as RealSimpleAriel on X to an audience of about 66,000, wrote five words into a thread:

I am long $MU

The post is public. One API call returns it as a long opened on MU, Micron Technology, at 952.735 USD, timestamped to the second, with the handle, the platform, the audience count and a link back to the message. The field set is documented if you want the response shape. Five words is the useful case rather than an unrepresentative one, because the extraction keys on the stated action rather than the prose around it.

The row does not stop at the entry, and that is what an entry feed cannot give you. In the same seven days the book recorded 276 adds and 96 trims alongside the openings, roughly 95% of each stated by the trader in a post. Whether somebody is still in a position, adding to it or reducing it is a different question from whether they ever opened one, and it is the question a product surface has to answer. Exits are the thin end, since most recorded closes are inferred from stale state rather than announced, so an exit here is a posted exit and never a sale.

If you are wiring any of this into a live product surface, the Social Signals Integration Playbook covers the integration patterns. It is gated, and written for product and engineering teams embedding signals in an app, so it is the wrong document for a research desk evaluating a dataset.

What each layer cannot do, stated as specifications

The vendors worth buying from attach their own doubt to the individual record instead of resolving it in a deck. Arkham does it with confidence badges on entity attribution, stating in its API documentation that "attribution is probabilistic, not binary claims". paste.trade calls its own performance figures derived observations rather than a realized track record.

Centaur does it with three flags on every event: whether the trader stated the action, whether the system inferred it, and whether the post described a position already held. That last flag ruins a leaderboard when it is missing, because a post saying somebody is already up on a trade will score beautifully and was never a call anyone could have acted on.

No position size. Asset, direction, entry price and timing exist. Notional, weight and allocation do not, and cannot be inferred from a post. Any figure for how much a trader committed would be invented. The honest unit is the return per settled call.

The coverage denominator is disclosure, not discussion. Forward Industries holds the largest Solana treasury of any listed company, and in the 90 days to 28 October 2026 a single tracked trader posted a position in it: one long, no shorts, with the API confirming the asset exists rather than failing a lookup. An attention product would have a great deal to say about that ticker. A positioning dataset has one row, and a thin answer is the correct answer from one. That is the clearest case for running two layers.

A post is a disclosure and not a fill. Centaur measures trades that traders disclosed publicly and does not verify broker or onchain activity. Where execution confirmation is the requirement, the layer is broker records or onchain settlement, and none of the four above substitutes for those.

Four questions, answered here first

Take these to any vendor in the set. Each has a different answer per category rather than a better or worse one, which is why they sort the four layers faster than a feature matrix does. Centaur's answers sit under each, because a buyer's checklist whose author will not sit for it is worth nothing.

1. What is one row, and what is its primary key? A project and a window, an asset and an interval, a document, or a position. The answer tells you what you can join it to. Here: a position, keyed on a position id, carrying an asset id, a trader id and the id of the message it came from.
2. Does a row carry a direction, and where did it come from? Stated by the author, tagged by the author, inferred by the vendor, or absent. If any of it is inferred, ask what share, and ask to see the flag on the row rather than a figure in a deck. Here: long or short on every row, with the three flags above, and the counts for each in the opening.
3. What share of the source material becomes a row? Every vendor in all four categories reads far more than it publishes. The ratio tells you what the extraction is doing, and whether a gap means the market went quiet or the pipeline missed it. Here: one stated opening per 177 messages read, from the figures at the top.
4. Was the price captured when the post went out, or attached afterwards? A price captured at publication time can be checked against the post. A price looked up later cannot, whatever the schema says. Here: captured at the timestamp of the message, with the message on the row, so any row can be opened and checked as the Micron one was.

The third is the cheapest to ask and the one most likely to go unanswered. None of the vendor pages read for this piece publishes a figure for it.

How to reproduce the figures above

Every number comes from a read of Centaur's partner API on 9 September 2026 covering the seven days to 08:06 UTC, except the MicroStrategy and Forward Industries figures, which come from a read on 28 October 2026 covering the previous 90 days. Stated openings filter to events the trader posted themselves, excluding entries inferred from context and posts describing a position already held. All of it moves, so treat any figure here as a reading taken on a date.

The stack most products actually want

Very few teams need one of these and none of the others, and the buying patterns show it. Moni lists Nansen among the teams it powers, an onchain analytics firm buying social data rather than building it. One tier up, among the houses a fund shortlists, Messari sells mindshare, The Tie and Context Analytics sell tweet-level scored sentiment and conversation volume, and LSEG's MarketPsych tags sentiment across news, social posts and transcripts.

Read against their own product pages, none of them sells a position event. Nansen and Arkham carry no social data at all, checkable against their published endpoint inventories, and on the pages checked neither do Kaiko, Amberdata or Coin Metrics. The layer is thinly occupied at every tier.

Which is why the decision is not about vendors. A trading front-end deciding what to put in front of a new user wants attention data to know what the market is discussing and positioning data to know which way credible participants are leaning, and the second is the harder half to assemble (the retention case is here). An agent wants narrative for context and attributed events so it can explain why it acted rather than assert a conclusion (more on that). A research desk wants the event table alone, because it intends to test the data rather than trust it.

Work out which of those you are, and how many of the four layers the thing you are shipping needs. If one of them is an attributed, source-linked position with a direction and a lifecycle, the partner documentation has the endpoints, the flags and the settled-outcome statuses.