TL;DR
- Promptwatch's ChatGPT numbers matched our own independent run within 4 points.
- It splits one cited page into several rows by URL parameter.
- Essential costs $95 a month, and content tools need $245.
- Its starter prompts name your brand, which inflates the score.
- Verdict: 4 out of 5, the best value tracker we have tested.
Promptwatch measures AI visibility accurately. We ran its prompts through ChatGPT ourselves and its numbers held. It also tracks more engines than its entry plan says it should, and it shows you the searches ChatGPT ran before it answered. Two things hold it back: its citation counts split one page into several rows, and its built-in agent could not find data the dashboard was showing.
We took the Essential plan and pointed it at HubSpot. Below is what we found, including one test it passed that most tools in this category have never been put through.
What Promptwatch is
Promptwatch is an AI search visibility tracker from Amsterdam, founded by Gijs de Groot and Klaas Foppen. It launched on 1 April 2025 and raised a €6 million seed round in July 2026, led by Seed + Speed Ventures. The company says more than 1,840 organizations use it, Duolingo, Fireflies and Monks among them. On G2 it holds 4.6 out of 5 from 20 reviews.
The core job is the same as every tool in this space. You give it questions your buyers might ask an AI assistant, it asks them on a schedule, and it reports how often your brand shows up, where, and next to whom. If the category is new to you, our guide to what AI visibility means covers the basics.
Promptwatch then adds a lot around that core. There are query fan-outs, a page tracker, a Prompt Explorer for finding questions, a chat agent, a sitemap crawler and a content generator. Further up the price list sit crawler logs, shopping tracking and an ads radar. The sidebar has nine sections and close to thirty screens, and part of this review is sorting out which of them the $95 plan actually opens.
One brand, 21 questions, seven engines
We used HubSpot as the test brand, the same one behind our Peec AI results, so the numbers line up across the series. Everything ran in English, from the United States.
Promptwatch wrote 10 prompts for us during onboarding. We added 11 more, all unbranded: 3 about email marketing and 3 about website and CMS software that we wrote ourselves, plus 5 about customer service that its AI wrote from our instruction. Then we grouped all 21 into five topics. Later we built a second monitor that asks our 16 unbranded questions on seven AI engines at once.
Finally, we sent the same 16 questions to ChatGPT ourselves, three times each, and counted how often HubSpot came up. That gave us 48 answers of our own to set against Promptwatch’s. The comparison is further down.
Pros
- Its brand detection matched a plain text search on all 37 ChatGPT answers we checked, and its HubSpot mention rate landed within 4 points of our own independent run
- Query fan-outs show the exact searches ChatGPT ran before answering, which explains a zero better than any score
- Topic grouping turns a flat prompt list into a map of where you win and where you are invisible
- Ran seven AI engines on the Essential plan, including Claude, Gemini, Copilot and Google AI Mode
- Shows the ad attached to each ChatGPT answer on the entry plan, even though the Ads Radar report is locked
- API and MCP access are included on the 95 dollar plan
Cons
- The same page gets split into several rows when AI answers add tracking parameters, so its citation count, and the Page Tracker built on it, run low
- The chat agent could not retrieve citations the dashboard was showing, and said the data was older than it was
- 3 of the 10 prompts it writes for you name your brand, which pushes the score up
- Content generation, the knowledge base, crawler logs, shopping and ads tracking all sit on higher tiers
- The pricing table says 4 models while its help center says every paid plan gets all of them, so get the cap confirmed in writing
- Perplexity answers stored in our test were short and named almost no brands, so treat that engine's numbers with care
What it builds for you before you start
The workspace arrives half-built. Promptwatch writes a brand book for HubSpot, picks eight competitors, invents three buyer personas, sets up three monitors, and writes ten prompts. One monitor is switched on. The other two sit inactive, aimed at the UK and Canada, ready to turn on.
Read those ten prompts before you trust the first score. Three of them name HubSpot directly: whether HubSpot is good for the full customer lifecycle, how to start on HubSpot for free, and whether HubSpot has AI tools built in. A brand-name question returns the brand nearly every time, and here those three scored 90% to 95%.
We saw the same thing with AthenaHQ, whose starter set also leaned branded. Scrunch went the other way and wrote questions that were harder on the brand than ours. Promptwatch labels these prompts “brand specific”, and you can filter them out. Most people will not, though, and the headline number includes them.
The competitor list has a similar blind spot. All eight picks were CRM or marketing-automation vendors: Salesforce, Zoho, Zendesk, Microsoft Dynamics 365, ActiveCampaign, Pipedrive, Freshworks and Keap. Nobody for email marketing alone, nobody for website builders. The fix is one tab away. A Suggestions list held 93 brands the engines had actually named, including Klaviyo, Mailchimp and Webflow. We promoted six of them in a minute. The list needs a quick clean, though: Beehiiv appeared twice, once on a misspelled domain, and Next.js was listed as a competitor.
Adding prompts is easy. You can write them yourself, upload a batch, generate them from keywords, or have the AI write them from an instruction. We asked for five help-desk questions with no brand names. They came back in about eight seconds, all unbranded, and editable before saving. Each prompt can carry its own language, from a list of 59 locales.
What the dashboard shows
Answers started arriving within minutes. A few hours in, across both monitors, HubSpot sat at 41.6% visibility with a 30.7% share of voice and an average position of #2.0. Salesforce came next at 17.7%, then Zendesk at 16.1% and Zoho at 15.2%. Sentiment averaged 75 out of 100.
Before the data lands, empty widgets fill with greyed-out sample brands, so you can see the layout before your numbers arrive. One thing to know: the visibility score is weighted by where you appear in the answer, so it is not a plain mention rate. A prompt where HubSpot was named in both answers, but low in the list, scored 40% to 45%.
Topics show where you are invisible
This is the screen that earned its place. Once prompts are grouped into topics, Promptwatch scores each topic and splits it by competitor. HubSpot’s headline 41.6% turned out to be an average of very different results.
| Topic | Prompts | HubSpot avg. visibility | Who leads instead |
|---|---|---|---|
| Sales and marketing alignment | 4 | 93% | HubSpot |
| Brand and competitors | 5 | 92% | HubSpot |
| Website and CMS | 3 | 89% | HubSpot |
| Customer service | 6 | 45% | Zendesk |
| Email marketing | 3 | 0% | ActiveCampaign |
HubSpot sells email marketing, and AI engines recommend it for everything around email except email. On customer service, where HubSpot has a whole product line, the competitor split put Zendesk at 48.4% and HubSpot at 10.9%. A single dashboard number hides both of those. This is why how you choose and group prompts matters more than which tool you buy.
The searches behind a zero
A score tells you that you are missing. Query fan-outs come closer to telling you why. ChatGPT runs its own web searches before it answers, and Promptwatch records them. It recorded 27 of them, all from ChatGPT.
Look at the email questions. For “what’s the best email marketing software for a small ecommerce business”, ChatGPT searched for “best email marketing software small ecommerce business klaviyo mailchimp omnisend shopify pricing features 2026”. It also ran site searches straight on klaviyo.com, mailchimp.com and omnisend.com. HubSpot was in none of them. The shortlist was settled before the search ran, and HubSpot never had a chance to be found.
Compare the CMS questions, where HubSpot scored 89%. There the searches included “hubspot” by name, next to Webflow and Contentful. In our own study, ChatGPT wrote brand names into its searches on 18 of 27 buying questions, so this is not a one-off. Promptwatch is the first tool in this series where you can watch it happen to your own brand, prompt by prompt, on the entry plan.
Seven engines on a four-engine plan
The pricing table says every brand plan tracks 4 models, and the tooltip names ChatGPT, Perplexity, Gemini and Claude. The model picker inside a monitor offers 18. Seven are tagged “Live”, meaning Promptwatch collects the answer from the real consumer app: ChatGPT, Google AI Overviews, Google AI Mode, Copilot, Gemini, Perplexity and Alexa. The other eleven, Claude among them, run through the model makers’ APIs.
We ticked seven on the Essential plan, and nothing stopped us. Five of them started answering almost at once. Google AI Overviews and AI Mode took a few hours longer, then filled in the same day. Promptwatch’s own help center settles it: “Every paid plan unlocks all active models.” So the pricing table undersells the plan. Get that confirmed in writing before you build your setup around more than four engines.
The engine heatmap is where the extra models pay off. On our 16 unbranded questions, one answer per engine, HubSpot’s visibility looked like this at the end of the first day:
| Engine | HubSpot visibility | Answers naming HubSpot |
|---|---|---|
| Claude | 59.1% | 12 of 16 |
| Google AI Overviews | 52.8% | 10 of 16 |
| Google AI Mode | 51.6% | 10 of 16 |
| ChatGPT | 47.6% | 9 of 16 |
| Gemini | 42.8% | 8 of 16 |
| Copilot | 38.9% | 8 of 16 |
| Perplexity | 7.5% | 2 of 16 |
The three email questions drew zero HubSpot mentions on all seven engines, so the email gap is not a ChatGPT quirk. The Google numbers also come with a caveat. We copied 13 of these questions into a separate Google-only monitor, and 11 of the 13 AI Overviews came back word for word the same, as did 11 of the 13 AI Mode answers. ChatGPT and Perplexity never repeated an answer across monitors. Google tends to serve the same AI answer for the same query, so a second monitor buys you no new Google data.
Claude favouring HubSpot fits what we found when we looked at Claude as a search engine. Perplexity is the number we would not act on yet. Every brand scored low there, not only HubSpot. On the home dashboard, which also counts the branded prompts, HubSpot’s Perplexity figure was 22.7%, so the question mix moves it a lot. The Perplexity answer we opened for the sales and marketing question was three generic sentences with no product named at all. That may be how Perplexity answered that day. It may also be how the answer was captured. Either way, we would wait for a few days of data before reading anything into it.
We checked its numbers against our own
A visibility tool is only worth paying for if its percentages are true. So we pulled all 37 ChatGPT answers Promptwatch had stored, then asked ChatGPT the same 16 unbranded questions ourselves through Bright Data’s ChatGPT scraper, three times each, from the US, on the same day.
First, detection. Promptwatch flags whether each answer mentions your brand. We ran a plain text search for “HubSpot” over the same 37 answers. The two agreed on all 37, with no misses and no false alarms.
Then the rates. On the 16 shared questions, Promptwatch’s ChatGPT answers named HubSpot 18 times out of 32, which is 56.2%. Ours named it 25 times out of 48, which is 52.1%. On a two-proportion test that gap is z = -0.37, well inside normal ChatGPT variation.
| Topic | Promptwatch | Our run |
|---|---|---|
| Email marketing | 0 of 6 | 0 of 9 |
| Website and CMS | 6 of 6 | 9 of 9 |
| Customer service | 4 of 12 | 4 of 18 |
| Sales and marketing alignment | 8 of 8 | 12 of 12 |
| All 16 prompts | 18 of 32 (56.2%) | 25 of 48 (52.1%) |
Every topic told the same story in both data sets. The zero on email marketing was a real zero, and the gap on customer service was real too. That is the result you want from a tracker. When we ran this check on Scrunch, it passed the same way.
With two answers per question, any single prompt can swing on one answer. Our citation variance study showed how much ChatGPT varies from one run to the next. Until a week of data has built up, read the topic and total figures.
Citations, and one page counted five times
The Sources section lists every URL the engines cited. In the first afternoon that was 630 unique URLs, each tagged by source type (corporate, blog, directory) and content type (product page, listicle, comparison, how-to). Each URL also carries its domain rating and its best and worst position. Open any single answer and you get the full text with inline citation chips, a citations tab, and the fan-out searches behind it.
We searched the list for HubSpot’s CRM product page and found it five times. The same page appeared as /products/crm?cb=1 with 5 citations, as /products/crm?+Nostalgia+Bait=undefined with 3, as the clean /products/crm with 3, as /products/crm?c1um1=7 with 3, and as /products/crm?gh_jid=6677219 with 1.
Those odd parameters come from the links inside the AI answers, so Promptwatch did not invent them. It just does not merge them. The page’s real total is 15 citations, but no single row says so. The citations engines attach are messy enough without the tool splitting them further.
The split carries straight into the Page Tracker, which is otherwise a good idea. You paste the URLs you care about, and it tells you how many prompts and responses cited each one, and when. It even fills in history right away. We added /products/crm and it showed 3 responses, which is the clean-URL row only. The other 12 citations were not counted. A HubSpot blog post that ChatGPT cited with a ?department= parameter showed as never cited at all.
If you report on cited pages, export the list and normalize the URLs yourself until this is fixed. It is the one place in our test where Promptwatch’s numbers were plainly wrong.
The Sources section also breaks out Reddit, YouTube and LinkedIn citations. In our first afternoon that was two Reddit threads, four LinkedIn posts and no YouTube videos. Each one comes with its upvotes or reactions and the position it was cited at. That matters more than it used to, given how far Reddit has fallen in ChatGPT’s citations.
The ads inside the answers
Promptwatch tags each stored answer with what it contains: tables, links, images, entities, ads. Of the 37 ChatGPT answers to our software questions, 26 carried an ad. Ads Radar, the report that adds these up by advertiser, needs the $579 Business plan. Each single answer’s Ads tab, though, opens on Essential.
On the Shopify email question, ChatGPT recommended Klaviyo, Omnisend and Shopify Email. Next to that answer sat an ad from Flodesk: “Looking for something easier?” That is a competitor buying its way into an answer that left it out, which is exactly how ChatGPT ads change the game for brands. One more detail: the answers we collected through Bright Data on the same day carried no ads. We do not know why the two collection methods differ.
Sentiment, with no reasons given
Sentiment is scored out of 100 for every answer that names you, then rolled up by prompt, topic and competitor. HubSpot averaged 74 to 75, labelled “moderately positive”. Branded prompts came in at 83, comparison prompts at 68.
The lowest score in our set was 40, on a Claude answer about missed support follow-ups. The only HubSpot line in that answer was “Already in HubSpot/Salesforce ecosystem: their native service hub modules”. That is a neutral passing mention. Nothing in the panel explains why it scored below the middle of the scale.
Treat a low score as a reason to open the answer and read it, as you would with any AI visibility metric that comes without its working.
Prompt Explorer finds questions, with some drift
Prompt Explorer suggests questions to track from five sources: your Search Console queries, your ranking keywords, a seed topic, Google’s People Also Ask, or a page on your site. Each suggestion comes with a volume and difficulty estimate and a one-click Track button.
We seeded it with “email marketing software”. It returned 10 prompts in about 20 seconds. One was about email software. The other nine drifted back to CRM, such as “what free CRM tools are actually worth using for a startup”. The seed lost out to what it knows about the brand. People Also Ask stayed on topic but did no filtering, so “Why is Outlook shutting down?” came back as a help-desk prompt worth tracking.
It is a decent starting list. Edit it before you track anything, especially when the topic you seed is one where your brand is weak, because that is exactly where the drift pulls hardest.
The agent, and what it got wrong
Promptwatch has a chat agent with access to your project. We asked why HubSpot was at 0% on email marketing, which sources the engines cited instead, and what to do about it. It thought for 16 seconds, sent off five research sub-tasks you can watch, and answered in about three minutes.
The brand half was right: Klaviyo, Mailchimp and ActiveCampaign lead the email questions, and HubSpot is absent. The citation half failed. The agent said it “could not pull the actual cited URLs for these prompts” because the responses “appear older than that or unlinked”. They were collected that morning, and the Sources page was showing omnisend.com, mailchimp.com and klaviyo.com among the most cited domains at that moment.
It flagged the gap instead of making up sources. Its three actions were reasonable: clean up tracking, publish a comparison page built around deliverability and Shopify segmentation, and get into the third-party lists that already mention Klaviyo and Mailchimp. It also waits for you to tick a checklist before it changes anything. It did suggest deactivating the prompts we had copied into the second monitor on purpose, which it took for duplicates.
That question cost 19 of the plan’s 500 monthly agent credits. At that rate Essential covers about two dozen real questions a month.
What Essential does not include
Click around the sidebar on the $95 plan and roughly a third of it opens an upgrade screen. Every locked screen says which plan unlocks it.
- Content Creation, Content Refresh and the Knowledge Base need Professional. The “Improve visibility” button on every weak prompt leads to a well-built content generator, with article types, length, persona, tone, sources and a content brief, and then to “0 of 0 content generations” on Essential.
- Content Agent, the scheduled plan-draft-publish workflow, needs Business.
- Crawler Logs, which name 24 AI crawlers in their preview, open an upgrade screen pointing to Professional.
- Shopping, Ads Radar and the audit log need Business.
- State and city targeting needs Business. Essential tracks by country.
- One project and one seat. A second brand or a colleague means upgrading.
Some parts are open on Essential but need your own website connected first. Those are Visitor Analytics, AI Insights, Topic Insights, Search Insights and Content Gap. We tracked HubSpot, a site we do not own, so we could not connect its Search Console or server logs, and we have not tested those screens. The sitemap crawler did work: it pulled 3,066 URLs from hubspot.com in about 15 seconds and began checking page speed, up to the plan’s 1,500-page limit. Crawler data is worth paying for if you act on it, and AI crawler logs show which pages bots actually read.
Pricing
| Essential | Professional | Business | |
|---|---|---|---|
| Monthly | $95 | $245 | $579 |
| Annual | $950 | $2,450 | $5,790 |
| Prompts | 50 | 150 | 350 |
| Responses a month | 6,000 | 18,000 | 42,000 |
| Projects | 1 | 2 | 5 |
| Agent credits a month | 500 | 1,500 | 2,500 |
| AI-written articles a month | None | 5 | 10 |
| Models (as listed) | 4 | 4 | 4 |
| Knowledge base | No | Yes | Yes |
| State and city tracking | No | No | Yes |
| Shopping and Ads Radar | No | No | Yes |
| API and MCP | Yes | Yes | Yes |
Annual billing is ten times the monthly price, so you get two months free. Agencies get a separate set of plans: Kick-off at $199, Growth at $399 and Scale at $799 a month. All three include unlimited projects and prompts, every model, and crawler logs, which makes Kick-off a better deal than Professional for anyone managing more than one brand.
Watch the response budget. Our first 10 prompts projected about 30 responses a day. After we added prompts and the seven-engine monitor, the projection rose to about 175 a day, which still fits inside 6,000 a month with little room left. Every engine you add multiplies the bill.
There is a 7-day trial, and it turns into a paid plan when it ends. Set a reminder if you are only evaluating.
Where it wins, where it falls short
It wins on accuracy and on explanation. The numbers matched our own run. Topics and fan-outs together tell you where you are missing and why: ChatGPT searched for Klaviyo by name before it answered. Per-answer ads, the engine heatmap and API access all come on the cheapest plan. For $95 that is more insight than we have seen from any tool at that price.
It falls short on data hygiene and on its own claims. Citations split by URL parameter, and the Page Tracker inherits the undercount. The agent misread the age of its own data. The plan table says 4 models while the product runs 7. The starter prompts lean branded. And a large share of the sidebar is a preview of higher tiers.
Against the field: Otterly.AI is cheaper for a basic monthly number, and Scrunch goes deeper on citation influence at more than twice the price. Promptwatch sits between them, nearer Scrunch on depth than on price. Our best AI visibility tools roundup compares the full field.
Should you use Promptwatch?
Promptwatch is the best value we have tested in AI visibility tracking. Its brand numbers are trustworthy, and its query fan-outs turn a bad score into something you can act on. Set it up with your own unbranded prompts grouped into topics, and it will show you the categories where you are invisible within a day.
Buy it if you want to understand why AI engines skip you, on a budget. Normalize the citation URLs yourself before you report on cited pages. Get the engine cap confirmed in writing before you rely on more than four. And skip it if what you really want is AI-written content, because that starts at $245.