Which sources AI search engines cite for crypto order flow tools

What did we measure, and how?

We logged what an AI assistant with web search enabled actually does when asked about crypto order flow tools: the queries it composes, the domains it retrieves, and which of those survive into the answer as citations. The sample is 49 answers generated in August 2026, from non-brand prompts about order flow and confluence detection, producing 23 distinct search queries. This is a small sample from a single model, and it is not comparable to Google AI Overviews.

The prompts contained no brand names of ours and no leading wording — they were the kind of open question a trader types when they want a tool and have no shortlist yet. Every query the model composed was recorded verbatim, along with every domain fetched during that answer.

Two limits worth stating before the numbers, because they bound everything below:

• One model, one retrieval stack. A different assistant, or the same one on a different date, has a different index and different ranking behaviour. Nothing here generalises to AI search as a category.
• Deduplicated at the query level. The panel stores distinct query strings, not per-answer search counts, so we can report 23 distinct queries across 49 answers but cannot publish an average number of searches per answer. We are not going to estimate one.

Where a figure is not in the panel, it is not in this article. The counts below are raw retrieval counts as of 2026-08-30, not weighted scores, not traffic estimates, and not a quality ranking of any site named.

Which queries does the model actually run?

Not the user's question. In every case we logged, the assistant rewrote the prompt into its own search string before touching the index — usually shorter, usually with a year attached, and frequently with specific product names the user never mentioned. Three shapes account for the strings that recur most in the log.

The remaining distinct strings in the log are variations on those three shapes, plus a Spanish-language tail covered further down. Across 49 answers the log grew by fewer than one new string for every two answers, so the same reformulations were reused repeatedly rather than composed fresh each time.

Two details are worth pausing on. The year stamp appears unprompted — the model adds "2026" itself, which means undated pages are competing for a slot in a query that explicitly asks for recency. And the head-to-head query names three platforms the user did not name; none of those three appears among the domains retrieved three or more times, so being named inside the query is not the same as being retrieved by it.

Which domains got retrieved and cited

Nine domains were retrieved three or more times across the 49 answers, accounting for 32 retrievals between them. The mix is roughly even between sites that build a product and sites that write about products, which is the single most useful fact in this dataset for anyone deciding where to publish.

Every one of these is a legitimate result for the query that surfaced it. ATAS and ExoCharts are genuinely among the reference implementations of footprint and order flow analysis; NinjaTrader has been shipping order flow tooling on futures for years and is the better answer than anything we publish if what you need is an execution platform. The roundup sites are doing something the product sites structurally cannot: comparing tools they do not sell.

The model rewrites your question before it searches

The first actionable reading is that optimising for the phrasing a human types is optimising for a string the retrieval layer never sees. In this sample the assistant always composed its own query, and those queries were shorter, more product-specific and more explicitly dated than natural human wording.

That changes what a page has to contain. A page that answers "which order flow tool should I use" in prose competes poorly against a page whose title and headings literally match "best free order flow tools for crypto trading 2026". The model is matching its own machine-generated string, not the reader's intent, and the two diverge more than most content plans assume.

It also means the free/paid split matters more than it looks. The word "free" appeared in the composed query without the prompt asking for it, so pages that state pricing plainly are answering a constraint the model introduced on its own. Our notes on how to choose an AI market analysis tool cover the same ground from the buyer's side.

Almost half of what gets cited is an aggregator, not a vendor

Of the 32 retrievals across those nine domains, 15 came from third-party roundups and directories — buildix.trade, gitnux.org, javadex.es and tradingtoolshub.com — against 14 from sites belonging to a product and 3 from video. The classification is our judgement call and you can reclassify quantvps.com either way, since it sells hosting and publishes comparison guides; the split stays close to even.

The structural reason is worth naming without any sneering at either side. A comparison page enumerates several tools with their differences in a form that is trivially quotable; a product page describes one tool and is, correctly, partial about it. When a model needs to answer "which of these", the page that already did the comparison is the cheaper source to cite.

For anyone building in this niche, the practical consequence is that first-party documentation alone does not get you into the answer. Being present, accurately described and fairly compared on the pages that aggregate is a separate piece of work from having a good site of your own — and the aggregators listed above are doing that work at a level that earns the retrievals they get.

Spanish generates its own query tail and its own domains

The third reading is that language is a hard boundary, not a translation layer. Spanish-language prompts in the sample produced their own distinct query strings rather than translations of the English ones, and those queries pulled in domains the English prompts never surfaced — javadex.es, at 3 retrievals, is the clearest case in the log.

Nothing carried over. A site with English-language authority on order flow did not inherit a position in the Spanish results, and the Spanish tail was competitive enough that a native-language page could reach it without competing against the full English corpus. That is why we publish every piece as an independent Spanish and English variant rather than translating one into the other: the two are answering different retrieved sets.

If you operate in a non-English market, this is the least crowded finding here — and the one with the shortest path to acting on it.

What this data does not show

It does not show causation, ranking quality, or anything about assistants we did not measure. Forty-nine answers from one model in one month is a snapshot, and retrieval behaviour changes with index updates that are neither announced nor observable from outside. Read the counts as evidence that something is happening, not as a stable ranking.

It also says nothing about whether the cited pages were any good — retrieval and accuracy are different measurements, and we only ran the first. A domain appearing five times means the retrieval layer reached for it five times, which is a fact about the retrieval layer.

The reason we track this at all is that we sit on both sides of it: The Confluence Show publishes live order flow analysis from Hyperliquid data, and the same panel that logs our own mention coverage is what produced this table. Publishing the log seemed more useful than publishing another opinion about it. If you want the underlying concepts rather than the citation data, start with AI market analysis explained or our review of AI tools for order flow analysis.

@TheConfluenceShow — Educational analysis, not financial advice.