What AI PR Actually Costs: 5 Questions to Ask Before You Buy

So I went looking at what happens after someone starts using RunPR.

RunPR writes journalist pitches using a rubric my cofounder Matt Pressberg built from years as a tech journalist and PR pro. He knows how to write a damn good pitch, by the way.

But a good pitch starts before the writing. It starts with current research, relevant sources and evidence a person can check. So I pulled apart one of RunPR's most research-heavy features: AI Visibility.

I expected the final answer to drive most of the cost. It didn't.

The expensive part happened before the answer showed up: searching the live web, checking sources, comparing results across AI engines and retrying when a provider failed.

That distinction matters if you work in PR. Two tools can use the same AI model and produce very different work. One might generate a quick answer from a prompt. Another might search for current evidence, compare several engines and show you enough detail to trust what you put in front of a client.

The model name on the homepage doesn't tell you which one you are buying.

The numbers, without the math soup

We measured a fixed 30-day window: 2026-07-18T15:04:07Z inclusive to 2026-08-17T15:04:07Z exclusive. Inside it, RunPR ran 74 production AI Visibility runs — 49 standard and 25 deep across six AI engines.

Those scans created 11,160 provider attempts. Of those, 10,448 returned a stored response and 712 stored a provider error. In plain English, an attempt means RunPR asked one AI engine one question. We counted both because failed work still uses time and can create a provider charge.

At the workload shape we observed, the modeled direct-variable API cost was:

• about $1.15 for a standard scan
• about $10.06 for a deep scan

These are historical modeled estimates for one RunPR workflow, built from invoice-reconciled, ledger-reconciled, modeled, and list-rate inputs. They are not an invoice-grade per-scan bill. The 57% figure below is a derived modeled allocation, not a line item on an invoice. They are not retail prices or a universal price list for AI PR software.

One engine accounted for about 57% of the estimated cost of a deep scan. It wasn't writing a response that was 57% better or longer. Each answer from that engine triggered about five requests and 14 billable web searches behind the scenes.

We asked for one answer. The provider did a small research project to produce it.

The takeaway for buyers

Compare the work behind the output. A useful AI PR tool should search current sources, show its evidence, explain what a run consumes and give a person control before anything goes out.

That is the standard we are building RunPR around.

What a PR agency is paying for

The paragraph at the end may be the cheapest step in an AI PR workflow. The useful work starts earlier:

• finding current coverage and sources
• checking whether a claim appears across more than one engine
• separating a useful citation from a random mention
• retrying a failed search without losing the whole job
• keeping enough evidence for a PR pro to review the result

A higher provider bill does not prove a tool is better. Providers can waste money too. Price and model choice tell you little without the workflow behind them.

For an agency, your team's time and credibility cost more than an API call. A cheap answer that sends an account lead to stale coverage or an irrelevant journalist can damage a client relationship. A thorough scan also wastes money if nobody can explain what decision it helped the team make.

Useful evidence per dollar is the measure that matters.

Five questions to ask before buying an AI PR tool

1. Does it search the live web?

Ask whether the tool checks current sources when you run it or relies on what the model already knows. That difference matters for recent coverage, journalist moves and fast-changing client news.

2. Can I see the evidence?

A PR professional should be able to open the source behind a claim. "The AI said so" will not survive a client call.

3. What happens after I click the button?

You do not need a tour of the vendor's infrastructure. You do need to know whether the tool searches, compares sources, checks more than one engine and handles failed results.

4. What does one credit or run buy?

Ask what counts as a standard run, what counts as a deep run and how much usage can vary. A bucket of mystery credits is hard to budget across client accounts.

5. Where does a person approve the work?

AI can gather and organize evidence. Your team owns the judgment. The product should make review easy and show how it reached the answer.

How RunPR answers those questions

RunPR searches current sources and captures the exact queries, cited sources, mentioned competitors and per-engine results behind each visibility scan. Standard scans cover ChatGPT, Perplexity and Exa. Deep scans add Claude, Gemini and Grok for a full six-engine read.

Our plan pages state the exact monthly allowance for each scan type, and add-on packs say exactly what they add. When the research turns into outreach, every pitch enters RunPR's approval queue. Nothing sends until a person reviews and approves it.

You can test the front end of that workflow on your own brand with the free visibility check. It takes about a minute and does not require an account.

What we changed inside RunPR

This audit changed how we track and control our own usage.

We now look at the full workflow instead of counting prompts alone. We separate measured provider charges from estimates, put limits around expensive scan behavior and check whether an engine adds evidence the others missed.

We also count internal and complimentary usage. I used to think of an internal scan as free because no customer was billed for it. The provider still sent us a bill. Well, of course it did.

Limitations

We are publishing these because they are part of the result.

Historical model usage coverage for this window was incomplete: 2,411 visibility rows against 11,160 stored attempts, with token columns unavailable. That is why the per-engine rates are modeled and invoice-reconciled allocations rather than invoice-grade per-run measurements.

This study covers one RunPR workflow during one 30-day period. The figures include direct variable API usage for AI Visibility scans. They do not include engineering time, human review, fixed infrastructure or every other part of RunPR.

Some rates came from provider invoices. Others used our best historical estimate because the old usage records did not contain enough detail to rebuild an exact bill for every scan. Provider pricing and behavior can change too.

For the provider-side accounting we used, xAI documents its cost tracking and management billing API publicly.

We are keeping those limits in the article because a precise-looking number can still be an estimate. PR people deal with that distinction every day. AI vendors should make it clear too.

Run a free AI visibility check to see where your brand appears, which competitors show up instead and which sources the engines cite. It takes about a minute and does not require an account.