Ask ten companies what an AI search agency does and you will get two answers that have almost nothing in common. Half describe a firm that builds a search engine into a product, so that a customer typing a half-formed question into a support box gets the right document back instead of a keyword match. The other half describe a firm that works on the outside of the business, trying to get a brand quoted when somebody asks ChatGPT or Google's AI Overviews for a recommendation. Both are real trades. Both are being sold under the same phrase. The budgets, the skills and the people who should own the project are different in every respect, and a buyer who confirms the category without confirming which trade they are buying loses a quarter finding out.
This is not a naming quibble. It decides who you brief, what a proposal should contain, and what counts as the work having succeeded. The distinction is easy to make once somebody points at it, and almost nobody does.
One phrase, two trades
The first trade is engineering. A firm in this business takes the content you already own, which might be a product catalogue, a documentation set, a claims archive or a decade of support tickets, and builds a retrieval system over it. The modern version of that work uses embeddings and a vector index so that a query matches on meaning rather than on the exact words, usually combined with conventional keyword ranking because pure semantic search is worse than people expect at names, part numbers and dates. The deliverable is a running system inside your product, and the people who judge it are your own engineers and your own users.
The second trade is closer to public relations than to software. A firm in this business studies how assistants assemble an answer, what they cite when they do, and what makes one source get quoted in place of another. The work that follows is content, structure, technical markup and the kind of third-party coverage that gives a model something to point at. Nothing ships into your product. What changes is whether your name appears when somebody asks a question you would like to be the answer to.
Firms doing the first sell to a head of product or a chief technology officer. Firms doing the second sell to a head of marketing. It is common for a company to hear one pitch, brief the other department, and spend eight weeks before anybody notices the mismatch.
| What you are comparing | Search built into your product | Visibility inside AI answers |
|---|---|---|
| Who is searching | Your existing customer, inside something you own | A stranger asking an assistant who to hire |
| Who should own it | Product or engineering | Marketing or communications |
| What gets delivered | A running retrieval system and its evaluation set | Published content, markup and third-party coverage |
| How you know it worked | A retrieval metric on a labelled question set | Citation share across a defined question list |
| Who controls the outcome | You do, because the system is yours | Partly the model vendor, which is why nothing is guaranteed |
Print that table and put it in front of whoever is about to sign. An AI search agency that fits the left-hand column is the wrong hire for the right-hand column, and the reverse is just as true.
Which one your problem actually needs
Before you brief an AI search agency, the question that settles it is short. Where does the search happen, and who is doing it?
- If the person searching is already your customer, already signed in, and searching inside something you own, the answer is the engineering trade. The failure you are fixing is that your own search is bad.
- If the person searching has never heard of you and is asking an assistant who to hire, the answer is the visibility trade. The failure you are fixing is that you are invisible at the moment of the question.
- If both are true, they are two projects with two budgets, and the engineering one almost always has the clearer payback, because it is measured against a system you control rather than against a model somebody else retrains without telling you.
Companies with a large owned corpus, which is to say anyone in insurance, legal services, manufacturing, healthcare administration or enterprise software, are usually buying the first and describing the second, because the phrase they have heard is the one that has been in the trade press. Companies selling a service to a small number of high-value buyers are usually buying the second and would waste money on the first.
What the engineering side quotes, and what it leaves out
A serious proposal for product-side search work covers five things, and a proposal that skips any of them is a proposal that has moved the hard part into a later invoice.
- Ingestion and chunking. How your documents are split matters more than the model. A contract split at the wrong boundary answers half a question with confidence.
- The retrieval design itself. Ask whether they are running hybrid retrieval and how they handle exact-match cases. A firm that says embeddings alone will do has not run this in production against real product codes.
- Evaluation. There should be a labelled question set and a number attached to it before the work starts, so that "better" is a measurement rather than a demo. This is the single most reliable thing to ask about, because the firms that do it bring it up unprompted.
- Latency and cost per query at your actual volume. Both are fine in a prototype and both are how these projects get cancelled in month five.
- Who operates it afterwards. A retrieval system drifts as the corpus grows. Somebody has to re-index, re-evaluate and re-tune, and that somebody is either on your payroll or on a retainer.
The same discipline applies to ownership, which is the part buyers forget to settle while they are still excited about the demo. Our own piece on what you actually own when an engagement ends goes through the questions in detail, and they apply to a search build more sharply than to most work, because the index, the evaluation set and the tuning are all things a firm can quite reasonably keep if nobody says otherwise.
What the visibility side can and cannot promise
The honest version of this trade is worth buying, and an AI search agency working on the visibility side can earn its fee. The dishonest version is everywhere, and it is recognisable by a single tell: a guaranteed placement. Nobody controls what a model says. The assistants change their retrieval behaviour on a schedule nobody outside the labs knows, and a firm that promises you a position is either misunderstanding the system or counting on you not to check.
What a good firm does promise is process. They will tell you which questions they are targeting, what sources currently get cited for those questions, what they intend to publish or fix, and how they will measure whether citation share moves. That is a reasonable engagement. Our guides on the difference between answer engine optimisation, generative engine optimisation and search engine optimisation and on getting recommended by AI assistants lay out what the work involves, and what to check before hiring an AEO agency covers the procurement side of it.
Research groups tracking the shift are worth reading before you commission anything. Stanford's Institute for Human-Centered Artificial Intelligence publishes the AI Index each year, and McKinsey's annual State of AI survey tracks where companies are actually putting this money rather than where vendors say they should.
The conversation that sorts firms in ten minutes
You do not need a technical background to put these to an AI search agency. Three questions do most of the work, and the answers separate the two trades immediately.
- "Does your deliverable run inside our product, or outside it?" A firm that cannot answer this in one sentence is selling you a category rather than a service.
- "What is the number that tells us this worked, and who measures it?" The engineering answer is a retrieval metric on a question set. The visibility answer is citation share on a defined question list. Anything else is a story.
- "Show me the last time this failed and what you changed." Every firm that has shipped either kind of work has one. The ones that do not have one have not shipped.
Ask those before you ask about price. A firm that answers all three cleanly is worth a rate conversation, and our guide to the questions to put to an AI agency before you hire carries the longer list for the second meeting.
Where to start looking
There is no single kind of AI search agency, and the directory splits the two trades apart rather than filing them under one heading. Firms building retrieval and search into products sit under AI search agencies and custom AI solutions, and the teams that handle the data plumbing underneath a search build sit under machine learning and data science. Browsing by the trade rather than by the phrase is the fastest way to avoid the mix-up this whole piece is about.
If you would rather describe the problem once than sort through categories, tell us what you are trying to build and we will point you at the agencies in our directory that do that specific kind of work. It costs nothing and there is no obligation to hire anyone.
Decide which of the two businesses you are buying before you take a single call. It is a ten-minute decision that saves a quarter, and the firms worth hiring will be relieved that you made it.
Sources
- Stanford Institute for Human-Centered Artificial Intelligence, AI Index Report, annual.
- McKinsey & Company, The State of AI, annual survey.