If you are choosing an AI tracking tool, you need a collection method that supports the comparisons your team intends to make. AI rank tracking products differ in collection, counting and reporting methods. Compare them using these six questions. Two of them concern details most vendors do not publish, and asking for them in writing is the fastest way to separate a measured product from a marketed one.
1. What does one unit of your plan buy?
Vendors sell prompts on plan, prompts per day, checks, credits, annual credits and calls. A plan named for a prompt count has not told you the engine count or the frequency, and those multiply.
Convert any quote into collections a month: prompts multiplied by engines multiplied by runs. The full conversion, with current vendor rates, is in how to compare AI search tool costs.
2. How many times is one prompt run per collection?
This is the most important question and the hardest to get answered.
Assistants return different answers to the same prompt at different times. If a product runs each prompt once per period and reports that single answer as a result, it is presenting one observation of a variable system as a measurement. If it runs each prompt several times and reports a rate, the result provides more evidence.
No competitor page read for this site publishes a sampling method. Ask directly, in writing, and ask what the reported figure represents: one answer, the most common answer, or a rate across runs.
The OppAlerts answer, for comparison: each prompt runs once per provider per scheduled collection and the report stores that one answer, not an aggregate. Read it as one observation in a series.
A vendor that cannot answer this clearly is selling a number whose reliability it has not characterized.
3. What happens when a collection fails?
Collections fail. Services rate-limit, change, and go down.
Two things to establish: whether a failed attempt consumes your allowance, and whether a gap in the data is visible in the report or silently filled with the previous value. A chart that quietly carries forward a stale reading is worse than one showing a gap, because it looks like evidence of stability.
In OppAlerts, credits are checked when the weekly assignment is saved rather than spent call by call, so a failed collection carries no separate charge and no separate refund.
4. Which engines, and what does adding one cost?
Engine coverage varies widely and is often a plan boundary rather than a setting. Some vendors track one engine on their entry plan and require a tier change for three. Others list every engine on every plan. One sells additional models as an add-on.
Write down the engines your reporting requires. Perplexity and Claude in particular are absent from several products, including OppAlerts, and that requirement alone eliminates options.
Then check whether the engine list you need is available at the volume you need, since engine coverage and collection volume often vary by tier.
5. Can you keep a series comparable?
The value of tracking is comparison over time, which only holds if the question stays fixed.
Ask how much history the plan retains, what happens to it on downgrade, and whether the vendor records when a prompt's wording was changed. Changing a prompt starts a new series because it asks a different question. If the report does not mark that change, its trend line combines two different measurements.
6. What does it let you do with the result?
A tracking report tells you a competitor is named where you are not, and that certain publications are cited. Neither is a task.
Ask what the product does next. Some generate content from the gaps. Some read your server logs for crawler activity. Some audit your site. Some, including this one, search for placements and return the contact route for them. Others stop at the report and expect you to have three other subscriptions.
That difference usually decides the purchase more than the engine count, and it is the part that is easiest to skip in a demo.
What no tracker can do
Three claims to treat as disqualifying if you hear them.
- Guaranteeing a position. There is usually no ordered list, answers vary between runs, and nobody controls the services being measured.
- Attributing traffic or revenue without connecting to your analytics. The path from an answer to a visit is frequently indirect or absent.
- Explaining why an assistant chose its sources. That mechanism is not observable from outside, and anything describing it is inference.
A short evaluation script
- Send the same defined workload to every vendor: prompt count, engine list, frequency. Ask what it costs on their meter.
- Ask the sampling question and the failure question in writing.
- Ask what history is retained and what happens on downgrade.
- Price it for your whole team, since some vendors charge per seat and some charge nothing.
- Bring ten of your own prompts to the trial rather than using the vendor's examples.
Before choosing a plan, ask how the service collects, counts and retains your results. You can then judge whether its reports will support your review process.
See AI rank tracking tools, why repeated answers vary, or how OppAlerts collects tracked prompts.