Ask a distributor what their search costs and you will get one number: the invoice from the search vendor. It is almost never the real number, and it is usually not even the biggest component.
This is an article about the other components — the ones that do not arrive as an invoice, and therefore never get compared when somebody is choosing between vendors. We are going to build a five-year model. We are not going to fill in the vendor prices for you, because we cannot: Algolia, Coveo and HawkSearch all quote custom, and any table on the internet claiming to know your price is guessing. What we can do is show you the shape of the bill, and where the parts hide.
We sell a one-time-licence search engine. An article arguing that recurring costs are underestimated is, structurally, an article arguing that you should buy from us. Check the arithmetic rather than trusting the conclusion — the model below works just as well for talking yourself into a subscription.
The five things a search stack costs
- The licence or subscription. The number everybody compares.
- Infrastructure. Zero for hosted SaaS. Substantial for self-hosted open source.
- Integration. One-off, but rarely small, and it recurs every time you replatform.
- Relevance maintenance. The invisible one. Somebody's time, every month, forever.
- The cost of bad results. The largest number on the list, and the only one nobody puts on the list.
1. The subscription, and why it grows
Hosted search is usually priced on records, on search operations, or on both. This is a rational way for a vendor to price, and it has an uncomfortable property for you: the bill goes up when your business does well.
Add a product line, the record count goes up. Run a successful campaign, the operation count goes up. Grow the catalog through an acquisition and you can cross a tier boundary without anyone making a decision. Public reviews of the major hosted platforms return to this point repeatedly — the pricing is described as difficult to forecast, and the tiers as a cliff rather than a slope.
Enterprise suites and distribution-specific platforms are typically annual subscriptions instead: a fixed number, renewed and renegotiated every year. Fixed is better than variable. It is still a number that arrives every year, and the direction of travel at renewal is well understood by everyone in the room.
The question is not what the subscription costs. It is what it costs in year five, at the catalog size you are planning for.
2. Infrastructure, and the Elasticsearch trap
Open-source search has no licence fee, which is why it looks free and why it so often is not.
Elasticsearch runs on the JVM, which consumes significant memory before doing any useful work. A production deployment is typically a multi-node cluster rather than a single machine, and that cluster needs sizing, monitoring, upgrading and occasionally rescuing at two in the morning. None of that arrives as an invoice from a search vendor. It arrives as cloud spend and as a fraction of an engineer.
Be honest about that fraction. If a competent engineer spends even one day a month on the search cluster — sizing, tuning, chasing a relevance complaint from the sales team — that is a meaningful annual cost at a fully loaded salary, and it is a cost that never ends.
3. Integration, and the replatform tax
Every search product has to be connected to two things: a catalog to index, and a storefront to answer queries. That is the integration, and for most stacks it is measured in days rather than months.
The part people forget is that it happens again. When you migrate commerce platforms — and distributors do, roughly once a decade — a search product that lives inside the platform as a plugin has to be re-integrated or replaced, and its merchandising configuration may not survive the trip. A search product that sits behind an API is insulated from that, because the storefront simply calls a different endpoint.
This is not a large number. It is a number that appears in the year you can least afford surprises.
4. Relevance maintenance, the cost nobody books
Here is where the model usually breaks.
In most search platforms, quality is not a property you buy. It is a state you maintain. Synonym dictionaries, boost rules, pinned results, redirects for known bad queries, fuzziness thresholds, analyzer configuration. Every one of those is somebody's ongoing work, and it does not stop, because the catalog does not stop.
New manufacturer, new naming conventions, new synonyms to write. New product line, new ranking rules. Sales complains that the wrong SKU shows first for a big customer's favourite query, so somebody pins it, and now there is one more rule in a list nobody has documented.
Ask your team a simple question: if the person who tuned our search left tomorrow, would anybody know how it was configured? The answer tells you whether you have bought a product or inherited a dependency.
5. The cost of bad results
This is the largest line in the model and it is almost never in the model, because it does not arrive as a bill. It arrives as an absence.
In our July 2026 field test of twenty distributor websites, several national names returned zero results for a single misspelled letter in "circuit breaker" — a product every one of them stocks by the thousand. That is not a relevance problem. That is a customer who arrived intending to spend money, was told the product does not exist, and left.
You can size this yourself, and you should, because it is the only number in this article that comes from your own data rather than a vendor's:
- Pull your site search logs for the last quarter.
- Filter to queries that returned zero results.
- Manually check a sample: how many of those were products you actually stock, misspelled or described rather than named?
- Multiply that share by your average order value and your search-to-order conversion rate.
Most distributors who run this exercise find a number that dwarfs the subscription they were arguing about. It is uncomfortable arithmetic, which is part of why it rarely gets done.
The five-year model
Fill this in with your own figures. The point is not our numbers — it is which cells are empty in the way you have been thinking about it.
| Cost line | Hosted SaaS | Self-hosted open source | One-time licence |
|---|---|---|---|
| Licence / subscription | Annual, grows with records & queries | None | Once |
| Infrastructure | None | Cluster, sized and monitored | One server you likely own |
| Integration | Once, plus on replatform | Once, plus on replatform | Once |
| Relevance maintenance | Rules, synonyms, boosts — ongoing | Analyzers, tuning, an engineer | No dictionaries to feed |
| Cost of zero-result searches | Depends entirely on configuration | Depends entirely on configuration | The design target |
| Year 5 direction | Higher than year 1 | Flat, plus salary inflation | Zero |
Where a one-time licence stops making sense
Since we sell one, here are the cases where the arithmetic goes the other way, and we would tell you so on a call.
- You have no servers and no intention of having any. If your policy is cloud-only and self-hosted software is not permitted, the recurring model is what you are buying, and that is a legitimate reason.
- Merchandising is a daily job. If a team lives inside a merchandising console pinning products and running campaigns, that console is the product. Ours does not exist yet. Theirs does.
- Your catalog is small and stable. Twenty thousand clean SKUs that rarely change is a problem most engines solve adequately. The cost difference is real, and it may not be worth a migration.
The profile where owning the software wins is narrower and more specific: a large, messy catalog that changes constantly, customers who type badly, a renewal that grows every year, and a server in a rack. That describes most distributors, which is why we build for them. It does not describe everyone.
The one number to take away
Before your next renewal, go and find out how many searches on your own site returned nothing last quarter, and how many of those were for products sitting in your own warehouse.
Whatever you decide about vendors, that number is yours, it is knowable, and almost nobody has it.