How to protect free queries in an AI search tool
Every free answer pays for search calls and a model before anyone subscribes. Why bulk research hides in normal questions, and which basic limits break.
Your customer of 14 months cancelled this morning, and the box that asks why is empty.
They paid $30 a month and used your AI search tool most working days. Their saved threads are full of competitor pricing and market sizing for clients, and the night before they left they exported every one. There's no ticket and no angry email, so you open their account to see what you can find.
Last Thursday at 4:20 p.m. in San Francisco, they asked what a client's rival charges. Your tool wrote a confident paragraph with two sources where there'd normally be six, and a small gray line underneath: "Some sources couldn't be loaded." They asked again and got the same line. They asked 14 questions between 4 and 5 p.m. that week, and 5 came back like that. Someone who pastes your citations into client briefs can't do much with two of them, so at some point they stopped asking you and went looking elsewhere.
This is a scenario, and every price in it is an assumption to replace with your own. But it's how AI search free queries abuse can reach you: through a customer who paid, and who never says what went wrong.
Start with what one answer costs.
- 3 search API calls at $0.005 each: $0.015
- 8 pages fetched at $0.001 each: $0.008
- 1 model call to read them and write the answer: $0.017
- Total: $0.04 an answer
Your free plan gives 10 answers a day. A typical free user asks about 12 questions a month, which costs you $0.48. An account that uses everything costs 10 × 30 × $0.04 = $12 a month, 25 times as much.
Now say someone has a spreadsheet of 6,000 companies and four questions about each: what they sell, who runs them, what they charge, when they last raised money. Here's an illustrative month of that.
Every line in that spreadsheet is something a real analyst might ask, and no account ever goes over its limit. If five people are doing this, it's 400 accounts and 120,000 answers, and 120,000 × $0.04 = $4,800 a month. That's $160 a day, more than 5 of your customers pay you in a whole month.
That money has already left. The search calls ran and the tokens were used, so you could delete all 400 accounts this afternoon and not get a cent back. Say you have 300 customers at $30, or $9,000 a month. Without those accounts you'd spend $7,000 and keep $2,000. With them you spend $11,800 and lose $2,800, and the $33,600 in your bank account now runs out in 12 months.
The bill isn't what cost you that customer, though. Your free answers reset at midnight UTC, which is 4 p.m. in San Francisco this time of year, and a script spends a daily allowance as soon as it comes back.
- 400 accounts × 10 answers × 3 searches = 12,000 search calls
- 12,000 calls ÷ 3,600 seconds = 3.3 calls a second for that first hour
- 3.3 + your usual 2.5 = 5.8 calls a second, on a plan that allows 5
In this illustrative afternoon, the bars are your search calls and the line is your plan's limit.
About 1 search call in 7 gets refused that hour (0.8 ÷ 5.8). With 3 searches behind each answer, roughly 37% of answers lose at least one (1 minus 6/7 × 6/7 × 6/7). That's your customer's 5 out of 14. Nothing went down, so nothing paged you.
Now you have to choose. You can buy a bigger search plan and pay more for the scripts to run faster. Or you can cut free answers, in a product people only trust after they've watched it handle something hard.
The scenario is made up. This isn't:
From our early beta
A team in our early beta woke up to a $6,000 OpenAI bill. Someone had created hundreds of free accounts and used up the complimentary AI credits on every one of them.
Why bulk research hides inside normal use
An AI video tool can at least see a stranger rendering 200 clips. You see a question about a company. Bulk research arrives as short, well-phrased questions, and someone preparing a pitch may well ask 40 of them in an afternoon.
Those answers are also worth far more than they cost you. Anyone can wire up a search API and a model and pay that $960 themselves, or get the same cited answers from you for free.
The obvious fix is a lower cap. Take it from 10 a day to 3, and the same spreadsheet needs 267 accounts instead of 80 (24,000 ÷ 90). Real free users, meanwhile, don't ask one question and leave. They read the answer and ask a follow-up, then another, and the fourth question is often where your tool pulls ahead of a normal search engine. A cap of 3 stops them right there. Setting limits that don't invite extra accounts is worth an afternoon, but a cap only counts what happens inside one account, and this is spread across 400 of them.
Limit the damage first
These steps won't stop extra accounts. They keep extra accounts from taking your customers down with them.
Price every mode. Quick answers, deep research and long follow-up threads cost different amounts, so work out what a free user costs for each one.
Separate free traffic from paid. Hold back part of your search rate limit and model budget for paying customers, then put a hard spending cap on free users. When free runs out, only free stops.
Spread out the resets. Give each account its own reset time, based on when it first asked, so allowances don't all come back in the same minute.
Cache what repeats. Popular questions can be answered once and reused. That does nothing for 24,000 different questions.
Make free answers cheaper. Use fewer searches and fewer pages, and keep deep research on paid plans. Read the results yourself first, because a thin first answer loses people too.
The basic checks, and how each one breaks
With a ceiling on the damage, most teams go after the accounts themselves, and email comes first. A unique-email rule compares text. Gmail ignores dots, and plus-tagged mail reaches the same inbox, so one mailbox can be written many ways.
| Basic fix | What it catches | What it misses | Who it wrongly blocks | What it costs to keep working |
|---|---|---|---|---|
| Email verification | Typos and dead addresses | Any real second inbox | People whose code lands in spam | Deliverability work |
| Disposable-domain blocks | Known throwaway services | New domains and ordinary mailboxes | People using privacy forwarding | A list someone updates forever |
| Gmail dots and plus addressing | One inbox written several ways | A second real mailbox | People who sort their mail with plus addresses | Upkeep as mail services change |
| Cookie limit on anonymous answers | A casual visitor coming back | Anyone with a cleared browser | Shared library and office computers | A number to keep retuning |
| CAPTCHA | Simple scripts | People signing up by hand | People using screen readers or privacy tools | Fees, and friction in front of a search box |
| Phone requirement | Casual second accounts | Anyone with another number | People without a reliable mobile number | A charge for every text |
| Card requirement | Casual second accounts | Anyone with another card | Students and journalists trying it first | Far fewer people ever see an answer |
| One per IP address | Repeat signups on one connection | Any other network | Research teams, newsrooms and campuses | Exceptions and tickets |
| VPN blocks | Known VPN ranges | Everything off your list | Analysts who search through one | Another list to maintain |
| Rate limits | Fast bursts from one account | 80 accounts asking 10 each | Real users deep in a research session | Tuning for each mode |
| Manual review | Accounts that look odd | Company questions, which look like research | Whoever looks odd to a tired reviewer | 400 accounts at 1 minute each is over 6 hours |
Manual review has a second problem in a search tool: it means reading what people asked, and your customers' research is often confidential.
Search also feels friction more than most products, because the whole product is a box you type into. Put a verified email, a puzzle, a phone number and a card in front of it, and a curious visitor leaves before seeing an answer. CAPTCHA has limits of its own.
So the rules become a second product, maintained by hand, with the search bill still running while you work on them.
Decide before the free answers go out
You don't have to build your own abuse prevention system. AbuseZero was built for the moment a new account claims your free answers, and it brings far more to that moment than any row in that table. Where your domain list goes stale, it checks mailbox aliases, disposable domains and mail routing. Where an IP block lands on a whole newsroom, it sees beyond an IP address with VPN, hosting, network reputation, Tor and proxy intelligence. When the addresses keep changing, it recognizes repeat devices across new accounts. Creating a fresh account isn't enough to erase previous activity. We built a proprietary algorithm for this, with AI that checks the claim again. AbuseZero flags abuse that single checks miss. Its multi-layer AI risk analysis examines 140,000+ signals and detects linked claims. And it blocks spammy bots, browser automation and malicious scripts while allowing genuine AI agents and assistants.
Your server makes one API call before the free allowance goes out and gets a clear allow or deny, under rules you control. Your app still counts the 10 answers a day and still delivers them. What changes is whether a new claim on that allowance goes through. Set eligibility to fit the allowance, use Observe mode to see what would be denied before anyone is blocked, and follow the quickstart to protect your first benefit.
Frequently asked questions
How many free queries should an AI search tool give?
Enough for one real research session, follow-ups included. Multiply that by your cost per answer and make sure you can afford it across every free user. This guide to free credit amounts walks through the numbers.
Should people have to sign up before their first answer?
Usually not, because the first answer is what sells a search tool. Keep anonymous answers few and cheap.
Can I recover what farmed answers cost?
Generally not, because your search and model providers charged for work that ran. You can still find the accounts that spent it.
Keep your search capacity for the people paying for it
Start with Builder
Pay as you go. Evaluate your first claim today.
See pricing
Builder, Pro, Scale and Enterprise.
Talk to us about Enterprise
High volumes and custom terms.