The API cost of querying nine engines is small. The cost that kills in-house builds is the two engines with no API, which require browser automation to reach at all.
At 100 prompts across the seven API-reachable engines, checked daily, the inference bill is roughly $64 to $127 a month. The two Google answer surfaces are not a bigger line on that bill. They are a different kind of problem.
Last updated: 30 August 2026. Model pricing and vendor terms change often, so this page is rechecked quarterly.
Build versus buy in this category is usually argued on the API bill, and the API bill is the one part that is genuinely cheap. That is why the argument keeps producing the wrong answer.
What does the API half actually cost?
Less than most teams assume. An answer to a tracked prompt is roughly 30 input tokens and 600 output tokens. Anthropic’s published pricing puts Claude Haiku 4.5 at $1 per million input and $5 per million output, and Claude Sonnet 5 at $2 and $10, which brackets the realistic range.
| Prompts | API-reachable engines | Runs per day | Answers per month | At Haiku rates | At Sonnet rates |
|---|---|---|---|---|---|
| 100 | 7 | 1 | 21,000 | $64 | $127 |
| 200 | 7 | 1 | 42,000 | $127 | $255 |
| 100 | 7 | 5 | 105,000 | $318 | $636 |
Computed from published per-token rates at 30 input and 600 output tokens per answer. Every provider prices differently, so treat this as the shape of the bill rather than the bill.
At the top row, the inference costs about as much as a team lunch. If that were the whole build, everyone would build.
Which engines can you actually reach?
Seven of nine have a public API you can call. Two do not, and they are both Google.
| Engine | How you reach it | Buildable? |
|---|---|---|
| ChatGPT | OpenAI API | Yes |
| Claude | Anthropic API | Yes |
| Perplexity | Perplexity API, with citations returned | Yes |
| Gemini | Google model API | Yes |
| DeepSeek, Grok, Meta AI | Provider APIs or hosted model endpoints | Yes |
| Google AI Overviews | A Search surface. No public API to query it | No |
| Google AI Mode | A Search surface. No public API to query it | No |
Perplexity’s API documentation is a good example of the buildable case: it returns “web-grounded answers with built-in citations in one call,” which is exactly the shape a tracker needs.
The two Google surfaces are a different category of thing. Google’s documentation on AI features describes them as places a page can appear, with eligibility rules, rather than as endpoints you can call. There is no key to request. The only way to see what an AI Overview said is to load a results page and read it.
Why are those two the thing that kills the build?
Because reading a results page at scale is a different engineering problem from calling an API, and it never stops being one.
- Why
- The answer renders in a page rather than arriving as JSON, so you are driving a browser and parsing a layout.
- Ongoing cost
- The layout changes without notice, and every change silently breaks your parser rather than erroring.
- Why
- Automated traffic at volume from one address does not behave like a person searching, and gets treated accordingly.
- Ongoing cost
- Proxy infrastructure, which is a recurring bill and an operational surface of its own.
- Why
- Automated access to a search product is governed by that product’s terms, not by an API agreement you signed.
- Ongoing cost
- A question for your legal team before you start, not a technical detail to resolve later.
- Why
- When collection fails for these two, it usually fails quietly rather than loudly.
- Ongoing cost
- A gap in your series that looks exactly like a drop in your visibility, which is the worst failure mode available.
Two of nine engines account for most of the engineering, all of the legal risk, and the only failure mode that lies to you.
What else does a real build have to carry?
Three costs that never appear in the first estimate, because they are maintenance rather than construction.
Model churn. OpenAI’s deprecations policy gives at least six months for generally available models, at least three for specialised variants, and warns that preview models “may be retired with much shorter notice, such as 2 weeks.” Multiply that across seven providers and migration becomes a standing item rather than an event.
Sampling discipline. One run per prompt is not a rate, so a build that queries once a day produces numbers you cannot defend. Doing it properly multiplies your bill by the run count, which is the arithmetic in enterprise ai visibility platform.
Answer storage and coding rules. Counting a brand mention is easy. Deciding consistently whether being named as the incumbent a challenger beats counts as a mention is a rule somebody has to write, apply identically every period, and defend later.
What about the half-build?
The most common outcome, and worth naming because teams arrive at it by accident rather than by choice. You build the seven API engines, skip the two Google surfaces, and tell yourself you will add them later.
That is a defensible product decision and a bad reporting one, unless you say so on every chart. A visibility number covering seven of nine engines is not comparable to anyone else’s covering nine, and it silently excludes the surfaces most likely to reach buyers who never leave a search page.
If you go this route, label it. Report per engine rather than as one blended figure, and state which surfaces are missing. A number that quietly omits Google is the kind of thing that gets discovered in a board meeting by somebody who read a competitor’s report.
So when does building actually make sense?
Three situations, and they have nothing to do with the API bill.
Note what is missing from that list. Saving money is not on it, because at the volumes most teams run, the inference is cheaper than the engineer-days spent maintaining the collector.
How should you actually price the decision?
In engineer-days rather than dollars, because the dollars are the small half and your day rate is the number you know.
Estimate three things: the days to build a collector for the seven API engines, the days to build and then permanently maintain browser automation for the two that have none, and the days per quarter absorbed by provider migrations. Add the inference figure from the table, which will be smaller than any of them.
Then compare that total against a subscription. If the two Google surfaces are in scope, the honest comparison is rarely close, and it gets less close every time a layout changes. If they are out of scope, build if you want to.
Whichever way it lands, the questions worth asking a vendor are the same ones you would have to answer yourself: how many times each prompt is queried, whether you get the raw answer text, and whether you can export it. Those are set out in best AI visibility tracking tools, and the sub-$100 options that make the build-versus-buy maths hardest are in cheap AI visibility tracking tool.