Bull AI Labs vs Profound: they log the visit, we test the fetch
Profound is the category's most mature AI answer-monitoring platform: $155M raised[01], 700+ enterprises[06], and Agent Analytics that reads your CDN logs to report which AI bots visited[03]. Bull AI Labs works one layer down. We fetch your pages live as the agents themselves - raw versus rendered, same URL, same second - detect what they can and cannot read, ship the engineering fix, and re-probe until it verifies. Pick Profound to observe the answer. Pick Bull AI Labs to fix the fetch.
By Sai Narendran · Updated 24 Aug 2026 · Sources dated at the foot
What Profound is genuinely great at.
Prompt Volumes is real conversation demand data, and nothing else in the category matches it. Profound has raised $155M[01], counts 700+ enterprises as customers per its published materials[06], holds SOC 2 Type II, and runs the most mature server-log crawler analytics in the category[03].
The log product is good. That is exactly why the distinction below matters: a log can only describe visits that already happened. It cannot tell you what the visitor could read.
The door camera.
“Agent Analytics is a door camera. It tells you a bot came to the door, when, and which badge it wore. What it cannot tell you is what happened when the door opened - what the bot could actually read once it got inside. We knock on the door ourselves. Ten times, wearing ten different badges, and we diff what comes back.”
Logs are retrospective. A bot has to have already visited. A log says nothing about a page no agent has hit yet.
Logs record status codes, not content. A 200 on an empty React shell is logged identically to a 200 on a complete page.
Logs cannot diff. What your browser assembles versus what ChatGPT-User actually received is the entire question, and no log contains it.
Three differences that decide what you can fix.
After the fact versus live.
Profound ingests CDN logs (Cloudflare, CloudFront, Vercel, Fastly) on a scheduled pipeline[03]. We probe live, on demand and on cadence, including pages no bot has ever hit.
Classification versus detection.
Profound classifies logged bots into training, search and agent classes, and does it well - Cloudflare adopted the same three-class taxonomy in July 2026[07]. We actively probe for blocks. In 802 persisted probe runs from our production corpus (17 Aug 2026), ChatGPT-User received a 403 on 96 runs. A browser user agent hitting the same properties received a 403 on 7[08]. Same sites, same paths - the difference is who is asking. A log-based product reads that as silence. A probe reads it as a finding with a fix.
Reporting versus the loop.
Their loop ends at a dashboard or a content draft. Ours ends at Probe, Persist, Generate, Ship, Re-probe, Verified. Two clocks govern the outcome, and we label them clearly: retrieval eligibility moves in 24 hours to 7 days; answer share takes 3 weeks to a quarter. We will not let anyone sell the second clock as the first.
These figures have decayed since the engagement ended, which we publish, because the decay is the argument for a standing instrument rather than a one-off report[10].
The gap widens on 15 September 2026.
From 15 September 2026, Cloudflare blocks Training and Agent crawlers by default on ad-displaying pages for new domains, while the Search class stays allowed[07]. Every week, more of the web 403s the exact user agents that answer buyer questions.
A log tells you after your citations went quiet. A probe tells you before.
Do not take our word for it. Watch the fetch.
Paste a URL and the two-fetch diff runs live: the same page fetched as a browser and as an AI agent, in the same second.
That was a demo domain. Run it on yours. Free, no signup, full diagnosis.
The research difference.
In May 2026 Ahrefs published the largest controlled study of schema markup and AI visibility: 1,885 treated pages against 4,000 matched controls, difference-in-differences[05]. All three numbers, including the two that mildly favour schema: AI Mode +2.4%, ChatGPT +2.2%, AI Overviews -4.6%. The first two are statistically indistinguishable from zero, and the one significant result points the wrong way.
So we capped schema at MEDIUM severity and 1 impact point and hard-capped it in code, so no generated sprint task can inflate it. We treat llms.txt the same way: retained as a check, zero citation weight.
As of 23 August 2026, the product screenshot in the main section of Profound's homepage shows a workflow template library, and one of the five visible templates is a Schema Markup Generator[09]. That is not a product defect. It is a category assumption. When a lever gets disproven, an inventory of workflow templates has every commercial reason to keep shipping it and a lab has every reason to downgrade it. We downgraded it.
One question to ask any vendor.
We hold ourselves to this first. In August 2026 we found our own ChatGPT runner was appending the brand name to the system message, inflating brand-found rank. We rebuilt detection so every engine receives the bare prompt and a mention only counts in the response body or citations, then re-ran the identical 50 prompts, 300 runs per engine. ChatGPT fell from 99% to 2%. Claude fell from 62% to 0.7%. Gemini was never injected and did not move, which is the control. We set a permanent measurement epoch and rewrote no historical rows[08].
Ask any vendor in this category, including us: “Does the prompt you send contain my brand name?” and “How many of my tracked prompts contain my own brand name?”
We have not tested anyone else's stack and we do not publish findings we have not run.
The full comparison.
| Criterion | Profound | Bull AI Labs |
|---|---|---|
| Answer monitoring | Yes. The most mature in the category, with screenshot auditability on captured answers. | Yes. Every tracked citation traces back to the fetch that produced it, so a drop gets a cause, not a shrug. |
| Prompt volume data | Yes. Real conversation demand data. Best in category, full stop. | No. You still leave with a tracked set: high-intent prompts generated from your audit and classified by query type. |
| Engine coverage | Deepest in the category. Up to 11 engines on Enterprise. | 9 engines per sprint cycle across agent, search, and training crawler classes, every score anchored to a persisted run. |
| Crawler analytics method | CDN log ingestion (Cloudflare, CloudFront, Vercel, Fastly) on a scheduled pipeline. Retrospective. | Live probes. 10 canonical user agents, raw versus rendered, on demand and on cadence, including pages no bot has visited yet. |
| Raw-vs-rendered diff | No. | Yes. Same URL, same second, browser render diffed against the agent fetch. You see exactly what AI search can cite. |
| Active 403 and edge-block detection | No. A block appears in logs only after a bot has already been turned away. | Yes. In 802 persisted probe runs (17 Aug 2026), ChatGPT-User drew 96 blocks against a browser agent’s 7. |
| Ships engineering fixes | No. The loop ends at a dashboard or a content draft. | Yes. Sprint tasks ship code-level fixes: render parity, crawl access, extraction structure. |
| Re-verifies shipped fixes | No. | Yes. A fix closes only when the re-probe confirms what the agent receives actually changed. |
| Published correction log | None that we have found. | Yes. A permanent measurement epoch, set August 2026, with zero historical rows rewritten. |
| Schema position | Ships a Schema Markup Generator as one of five visible homepage workflow templates (23 Aug 2026). | Capped at MEDIUM severity and 1 impact point, hard-capped in code after the Ahrefs May 2026 data. |
| Entry price | $99/mo billed yearly. ChatGPT only, 50 prompts, 1 seat, no exports, no API. | Free. The full diagnosis - two-fetch diff, competitor citations, engineering findings - before you pay anything. |
| Free tier | No free trial documented in 2026 reviews. | Yes. Proof is never gated. Only remedy is paid. |
What each dollar buys.
Profound
Public pricing since 11 Jun 2026[02]
- Starter, $99/mo billed yearly. ChatGPT only, 50 prompts, 1 seat, no exports, no API.
- Growth, $399/mo. 3 engines, 100 prompts, CSV/JSON export.
- Enterprise, custom. Up to 11 engines, SSO/SAML, API, Prompt Volumes.
Documented 2026 reviews note the API is Enterprise-only and that no free trial is offered. Fair to hold both against them.
Bull AI Labs
The audit is free
The free audit shows the full diagnosis: the two-fetch diff on your own pages, competitor citations on your queries, and every engineering finding behind them. Paid plans start when you want the remedy: shipped fixes, continuous probes, and re-verification on cadence.
Never gate proof. Only gate remedy.
Profound charges $399/mo before you learn anything. Our full diagnosis is free. Never gate proof. Only gate remedy.
Which one is yours.
Pick Profound if
- You need prompt-volume demand data at Fortune 500 planning scale.
- You need SOC 2 Type II for procurement today, with no exception process.
- You have an in-house team that only wants observation, not shipped fixes.
Pick Bull AI Labs if
- Your site runs on React, Next, or Vue and you have never seen your raw fetch.
- You are on Cloudflare and have never tested whether ChatGPT-User gets a 403.
- Your citations dropped and your dashboard cannot say why.
- You want fixes shipped and verified, not recommended.
- You want proof before you pay a dollar.
Straight answers.
Profound observes more answers than we do. Observing answers was never the bottleneck; fetching pages is.
We do not have SOC 2 Type II yet. The audit runs against your public pages and needs no data access.
Our probes run from one region today, disclosed on every report.
Questions buyers actually ask.
Does Profound detect blocked AI crawlers?
Profound's Agent Analytics reads your CDN logs and reports which AI bots visited, including the status codes they received. It can surface a 403 after the fact, but only for bots that already visited. It does not actively probe your pages to find blocks before any bot arrives.
Does Profound show what AI can read on my page?
No. Profound reports which bots visited and how your brand appears in captured answers. It does not fetch your page the way an AI agent fetches it, so it cannot show what the agent could actually read. That raw versus rendered comparison is the layer Bull AI Labs is built on.
What is a two-fetch diff?
A two-fetch diff loads the same URL twice in the same second: once as a browser that executes JavaScript, once as an AI agent that does not. Comparing the two responses shows exactly which content AI search can cite and which content never arrives. Bull AI Labs probes ten crawler user agents this way.
What is the difference between Bull AI Labs and Profound?
Profound observes AI answers and crawler visits after they happen, through answer monitoring and CDN log analytics. Bull AI Labs tests the live fetch itself, detects what agents can actually read on your pages, ships the engineering fix, and re-probes until the fix verifies. Observation layer versus engineering layer.
Is Profound worth $399 a month?
For enterprise teams that need prompt-volume demand data, the deepest engine coverage, and board-grade answer-share reporting, the Growth plan can justify its price. Teams that mainly need to know what AI agents can read on their site can start with the free Bull AI Labs audit before paying anything.
Does schema markup improve AI citations?
Not measurably, on current evidence. Ahrefs studied 1,885 treated pages against 4,000 matched controls in May 2026: AI Mode plus 2.4 percent, ChatGPT plus 2.2 percent, AI Overviews minus 4.6 percent. The first two are statistically indistinguishable from zero. Bull AI Labs caps schema at one impact point.
Do AI assistants run JavaScript?
Generally no. Agent-class crawlers such as ChatGPT-User, Claude-User, and Perplexity-User fetch raw HTML and do not execute JavaScript. Vercel measured 569 million GPTBot and 370 million ClaudeBot requests in a single month with zero JavaScript execution. Googlebot is the notable exception: it renders.
Can I test this without paying?
Yes. The Bull AI Labs audit is free and shows the full diagnosis: the two-fetch diff on your own pages, competitor citations on your queries, and the engineering findings behind them. Proof is never gated. Only the remedy, shipped fixes and continuous re-probes, is paid.
Profound can tell you a bot visited. Find out what it could read.
Sources · numbered and dated
- 01 ·Fortune, 24 Feb 2026. Profound $96M Series C at a $1B valuation led by Lightspeed; $155M raised in total.
- 02 ·Profound public pricing page, verified 24 Aug 2026. Starter $99/mo billed yearly, Growth $399/mo, Enterprise custom.
- 03 ·Profound Agent Analytics integration docs. CDN log ingestion across Cloudflare, CloudFront, Vercel, and Fastly.
- 04 ·Vercel AI crawler study: 569M GPTBot and 370M ClaudeBot requests in one month, zero JavaScript execution. OpenAI crawler documentation change, 9 Dec 2025.
- 05 ·Ahrefs, 11 May 2026. Schema study, n=1,885 treated pages vs 4,000 matched controls, difference-in-differences.
- 06 ·Profound's published materials, 2026. 700+ enterprises, including more than 10% of the Fortune 500.
- 07 ·Cloudflare. 1 Jul 2025: default AI bot blocking. July 2026: three-class crawler taxonomy update, and from 15 Sep 2026 Training and Agent crawlers blocked by default on ad-displaying pages for new domains while Search stays allowed.
- 08 ·Bull AI Labs production corpus. 802 persisted probe runs, 17 Aug 2026. Measurement-epoch correction, August 2026: no historical rows rewritten.
- 09 ·Profound homepage product screenshot, 23 Aug 2026, 12:57 PT. Screenshot on file. Schema Markup Generator appears as one of five visible workflow templates.
- 10 ·Freeletics 12-week sprint engagement. Figures shown as measured at engagement end; they have decayed since, and the decay is noted wherever the figures appear.