Buyer questions, AI infrastructure
The questions that decide which vector database gets recommended.
Four kinds of question decide an AI infrastructure category, and three of them never carry a brand name. This is what each one looks like, what the model assembles the answer from, and who tends to hold it today. Everything here is runnable without us.
One rung in four carries a brand name. The other three decide the category without one, which is why a branded search report reads clean while the category is being lost.
76 per 100
results in this category are listicle-shaped
Our own measurement. The densest answer surface we have measured.
82 hosts
carried the citations we counted
901 citations across 60 answers. There is no single site to buy onto.
7 of 12
queries flipped outcome across identical repeats
One model, five repeats each. A single read is a coin flip.
The short version
One definition, in a paragraph a model can quote.
Definition first, under sixty words, no pronoun pointing at anything outside itself. It is the same shape we build for clients, on our own page, which is the only honest way to sell it.
Short answer
What are the money queries for an AI infrastructure company?
A money query is a question an AI infrastructure buyer asks an assistant with a budget behind it. Money queries fall into four kinds: the category question, the head to head, the constraint question, and the task question. Three of the four carry no brand name, and third-party pages decide all four.
bestIntent { your category }Category for productionConstraint no brandBrand
Everything a buyer needed to decide is in the string, and none of it is your name. The answer supplies the missing part, which is the part you are selling.
The four rungs
Four kinds of question, and three of them never say your name.
Sorted by how much of the evaluation each one decides, not by how much volume a keyword tool reports for it. Volume and purchase intent come apart badly in this category.
The category question
No brand name
- best vector database for rag
- open source agent framework comparison
- llm observability tools compared
- What decides it
- Roundups, awesome-lists and benchmark writeups. The model assembles a shortlist out of pages that already rank several tools against each other, so the real question is which of those pages name you.
- Who holds it today
- The incumbent, on nearly every run. This is the rung a category definer already holds, and the one a challenger has to be added to rather than argued into.
The head to head
Carries a brand name
- yourstack.example alternatives
- yourstack.example vs incumbent.example
- is a managed vector store worth it for production rag
- What decides it
- Comparison pages, review sites and the threads where somebody has run both. A brand name in the question pulls in whatever has been written about that brand, including by people with nothing to gain from it.
- Who holds it today
- Whoever the question names, plus whoever the comparison pages set opposite them. Being the named alternative on somebody else's page is a real position and it is the most winnable rung here.
The constraint question
No brand name
- cheapest embedding api at scale
- self hosted vector database
- vector database with metadata filtering
- What decides it
- Docs and pricing pages, read as facts rather than as marketing. A constraint question is answered by whichever page states the number plainly enough to be quoted verbatim.
- Who holds it today
- Frequently nobody in particular. Constraint answers drift toward whichever page is easiest to extract, which makes this the cheapest rung on the ladder to go and take.
The task question
No brand name
- how do i chunk documents for retrieval
- why is my rag returning stale answers
- how to evaluate a retrieval pipeline
- What decides it
- Community threads, issue trackers and reference material a model can lift a working example out of.
- Who holds it today
- Usually unclaimed. It carries the most volume and the least brand intent, and it is where a tool gets recommended inside an answer to a question that was never about tools.
Only the second rung carries a brand name, and on your category it is usually somebody else's. The other three settle who gets recommended without your name being typed once.
One query, all the way through
The answer is written before your buyer reaches your site.
One question from the first rung, taken end to end. The query, the answer it produces, and the pages the answer was assembled from.
Your buyer asks
The answer they get
For production RAG, most teams land on Incumbent DB1. It pairs hybrid search with metadata filtering at scale.2
Where the citations resolve
reddit.com38%
g2.com26%
news.ycombinator.com21%- The query
- One question from the first rung, with no brand name in it
- The answer
- Names one product, and gives the reader no second option
- The hosts
- Three third-party pages carried the citations behind it
- Your domain
- Present in the run, cited by none of them
Run it yourself
Five minutes settles whether any of this is worth paying for.
Five minutes, no tooling, and it settles whether there is anything here worth paying for. Run it before you talk to us, and run it before you talk to anybody else in this category.
- 01
Pick the rung
Take one category question from the first rung above. No brand name in it, yours or anybody else's.
- 02
Ask it five times
Same wording, same engine, five separate sessions. Repeats are the method rather than a detail, because a single read flips.
- 03
Change engine
Repeat the five on a second assistant. Answers assembled from different retrieval sets disagree, and the disagreement is information.
- 04
Count the mentions
Write down how many of the ten runs named you at all, and which names came back instead.
Named in all ten
You are the definer. The category answer and the head-to-head answer already return you, so there is nothing here to buy. Keep your money.
Named in some
You are in the answer set and not reliably. This is the position the work is built for, because the gap is real and it is measurable.
Named in none
Your category is being decided without you. The first job is diagnosis rather than volume, because publishing more pages does not fix a retrieval problem.
one category question, same model, same day
Nothing was altered between calls. A single reading of where you stand is a coin flip wearing the clothes of a metric, which is why the method counts repeats instead.
A perfectly on-vertical category definer is a bad engagement for both sides. A slightly off-vertical challenger is a good one.
The method, given away
How to build the list without hiring anyone.
The list is the deliverable, and building one is not proprietary. This is the same sequence we run, written out so you can run it yourself.
26 of 1,240 survive. The 1,214 that do not are what a keyword export is, and running the month against them is the ordinary way this budget gets spent.
- 01
Disambiguate
Write each question the way a person types it, not the way a keyword tool reports it. One question per line, no operators, no truncation.
- 02
Strip the brand
Remove your own name from every line you can. A question carrying your name is one you already win, and it only fires for somebody who has heard of you somewhere else.
- 03
Ask for the budget
Keep a line only if somebody asking it could reasonably buy something within the week. That single test removes most of a keyword export.
- 04
Repeat five times
Ask each surviving question five times on the same engine and write down every answer. We ran twelve queries five times each and seven flipped outcome, so one read decides nothing.
- 05
Count, do not score
Record how many of the five runs named you rather than a single number. The count is what a control group can later be compared against; a score is not.
What counts
Most of a keyword export is not a money query.
Most of what a keyword export returns does not belong on this list. The cuts are what make the remaining twenty or thirty questions mean something.
yourstack.example alternatives
yourstack.example1 is the lighter of the two and covers the same core case.
Named, and you already win this
best vector database for rag
Three options come up most often: the incumbent1, a second name, and an open source project.2
yourstack.example does not appear
Same company, same day. The left panel is the one a branded report shows you, and the right panel is the one the budget is decided on.
Keep it if
- A category question with no brand name, where an answer returns a shortlist.
- A head to head that names a competitor rather than you.
- A constraint question whose answer is a number somebody has to state.
- A task question your product genuinely resolves, asked by somebody mid-build.
Cut it if
- Your own product name. You already win it, and it only fires after somebody has heard of you somewhere else.
- A head term from a keyword tool. Volume without purchase intent produces a list that moves and sells nothing.
- Anything a search of your docs answers faster than an assistant does.
- A question with no budget behind it, however often it is asked.
What the list becomes
A validated list is a baseline, not a report.
A finished list is a baseline rather than a report. Once the questions are validated, the set splits, gets written down, and becomes the thing a later claim is measured against.
- Validated
- 5 identical repeats before one query counts
- Split
- A treated arm and a held-out control arm
- Registered
- Both written down before any work starts
- Depth
- 40 queries by 40 samples, 9.9pp floor
- Readout
- Difference-in-differences at day 90
Stamped on the divide
The list is worth nothing until something is committed to against it. Two of these three are done before any work starts, and the third is the only one that can surprise anybody.
Citation sets churn, so a level reading taken twice tells you almost nothing. Ahrefs measured across 43,000 keywords that AI Overviews persist 2.15 days and 45.5 percent of citations change between consecutive observations, while the answer stays 95 percent semantically identical. A held-out control is what survives that.
Start here, free
We build the list for you: 12 queries, 5 repeats each, 3 engines.
Delivered in five working days, and walked through live rather than emailed as a PDF. You get your questions built and disambiguated, your failure mode named, the hosts cited on those questions ranked by citation depth, and a pre-registered baseline you can hold anyone to afterwards, including us.
The rest of the cluster
One play, run in one niche, all the way down.
One method, three ways in
Questions
The five we get asked about building the list.
Not covered here? The Gap Report costs nothing and answers most of the rest with your own data.
How many money queries should an AI infrastructure company track?
Twenty to thirty across one cluster is the working set for a single category, which is what the entry engagement covers. Sixty to ninety across three clusters is the depth at which the instrument reads reliably, because the minimum detectable lift at 40 queries by 40 samples is 9.9 percentage points and thinner designs cannot see a real win.
Why does the category question matter more than our brand query?
Your brand query is the one question you already win, and it only fires for a buyer who has heard of you somewhere else. The category question is the top of the funnel: it returns a shortlist assembled from third-party pages, and the shortlist is the evaluation.
Can we just export these from a keyword tool?
A keyword tool reports what people type into a search box, ranked by volume. A money query is defined by purchase intent rather than by volume, and the task rung in particular is phrased as a sentence that no keyword report surfaces. The export is a starting point that needs the four cuts above applied to it.
The answers change every time we ask. Is this measurable at all?
The level is not measurable reliably. We ran twelve queries five times each on one model and seven flipped outcome. The difference between a treated arm and a matched control still measures cleanly, because unbiased noise cancels: across 20,000 random splits with no work applied, the null difference centred on zero.
What is in the free Gap Report?
Twelve money queries, five repeats each, across three engines, delivered in five working days. It names which of the nine failure modes you are in, ranks the hosts cited on your queries by citation depth, and sets a pre-registered baseline. It is walked through on a call rather than emailed as a PDF.
Free gap report
When the model answers,
be the one it names
Tell us your category and the questions your buyers ask. We run them against live AI answers and walk you through what came back. If you are already winning, we will tell you that too.
12 queries · 5 repeats each · 3 engines · 5 working days
Free · Walked through live · Five working days
Queries we run
Cited instead of you
reddit.com38%
g2.com26%
news.ycombinator.com21%91.5% of citations point off-site