Best AI visibility tools: two machines sold under one name

What are the best AI visibility tools, and what does each one actually measure?

Set the criteria before looking at any product, because criteria chosen after the fact are chosen to be won. Five questions decide which tool fits, and none of them is about feature count.

The first is what the tool observes. A tool either watches what models say, or it watches what your server serves. Those are different systems with different failure modes, and no product observes both halves equally well. The second is whether the number can be checked twice. Ask a model the same question two days apart and the answer can differ with nothing on your site having changed, because the model, its index and its retrieval step all moved underneath you. Fetch your own homepage twice and the bytes are identical unless you deployed. The third is what the tool needs from you: a public URL, a login to your analytics, or your raw server logs. The fourth is the published price and what unit it is charged in. The fifth is the one buyers ask last and should ask first, which is what the tool cannot tell you at all.

Run those five over the category and it splits cleanly in two. Answer trackers ask models a panel of prompts on a cadence and diff the results into mentions, share of voice and citation counts. They are the larger half of the market by vendor count and the one most people mean by generative engine optimization tooling. Crawler access tools sit on the other side of the same pipeline and ask whether an AI crawler can reach and parse the page before any of that happens. The order matters, because the second is a precondition for the first: a page no crawler can read cannot be cited by a system that grounds its answers in retrieved pages, however many prompts you track. What that looks like when you actually run it is on what AI says about you, which is the answer tracking side of this product rather than the scanning side.

The honest summary of the split is that answer trackers measure an outcome you do not control, and crawler access tools measure an input you do. Neither is the better purchase in the abstract. If your pages are already readable and you need to know which of six brands a model names, the tracker is the correct buy and a scanner will bore you. If your pages are invisible to a fetch, the tracker will faithfully report that nothing mentions you, month after month, without ever telling you why.

Answer trackers: what the published plans actually buy

Four vendors in this half publish enough on a public page to quote. All four were read on 1 September 2026 and none of them was logged into, so everything below is what a prospective buyer sees before paying.

Profound's pricing page published three tiers. Starter is 99 dollars a month billed yearly, covering ChatGPT as the only answer engine, with 50 prompts tracked, 100 agents credits a month and 1 seat. Growth is 399 dollars a month billed yearly, labelled as the popular option, with 3 answer engines tracked, 100 prompts tracked, 400 agents credits and 3 seats. Enterprise is custom, and the page states up to 9 answer engines tracked with a tailored prompt tracking plan, alongside SSO, SAML and SOC 2. Both paid tiers are advertised as billed yearly with two months free. Our longer write up sits at our page on Profound.

Semrush sells this as an AI Visibility Toolkit rather than as a standalone product, and its pricing page published a Base plan at 99 dollars a month per domain billed annually. That plan states tracking for 25 custom prompts with daily AI rankings, one domain for brand performance analysis, mentions drawn from ChatGPT, Google AI, Gemini and Perplexity, a site audit for AI readiness, and 300 reports per day. Add ons are listed separately, including additional users starting at 45 dollars a month. The distinction worth holding is that Semrush is the only product in this group whose published plan covers both halves of the split above in one subscription, which is what an incumbent distribution channel buys you. Our page on Semrush covers the rest.

Two vendors published less than the others on the day we looked, and saying so precisely matters more than filling the cell. Peec AI's pricing page published four tier names, Starter, Pro, Advanced and Enterprise, each described by who it is for, with a feature grid naming AI shopping, number of models included, projects, countries per project and tracking cadence. The amounts were not present in the page content we could read, so no Peec price appears in this post; our page on Peec is where we keep what we do know. Otterly's pricing page returned no readable content to us at all on 1 September 2026, so no Otterly figure is quoted anywhere here, and our page on Otterly carries the older research instead. A tool that cannot read a page says so rather than guessing at it, which is the same standard this product is held to when a scan degrades.

None of those four links point at the vendors themselves, and that is deliberate rather than an oversight: this site does not pass link authority to companies it competes with, so rivals are named in prose and the reader is sent to our own comparison at the comparison index. If you want to verify a figure above, the pricing pages are at tryprofound.com/pricing, semrush.com/pricing/ai, peec.ai/pricing and otterly.ai/pricing, written as plain text so you can go and check us.

Why a prompt count is not a unit you can rank on

Every listicle in this category eventually builds a table with a prompts column and sorts it. That column is three different measurements wearing one label, and the differences are large enough to reverse the ordering.

Profound publishes prompts tracked, which is a size of panel held under observation. Semrush publishes 25 custom prompts with daily AI rankings, which pairs a panel size with a cadence in the same sentence. Lantad's PLAN_LIMITS in core/src/config.ts, read on 1 September 2026, records prompts per run and keeps the cadence in a separate field called reruns per week: the free tier is 5 prompts per run with 0 paid engines and 0 reruns, Starter is 10 prompts per run with 3 paid engines and 1 rerun a week, Pro is 15 prompts with 5 paid engines and 1 rerun, and Business is 25 prompts with 5 paid engines and 2 reruns. Those are settings that somebody chose against a cost model, not observations about anything, and the accompanying PLAN_PRICING constant records the published prices of 59, 199 and 399 US dollars a month for those three paid tiers.

Now put the two 399 dollar plans beside each other, because this is the comparison that does not flatter us. Profound Growth at 399 dollars a month billed yearly publishes 100 prompts tracked across 3 answer engines. Lantad Business at 399 dollars a month sets 25 prompts per run across 5 paid engines, rerun twice a week. On the axis a buyer of an answer tracker actually cares about, the number of distinct questions under observation, the rival publishes four times what we do at the same price, and a buyer whose whole job is panel breadth should read that and choose accordingly. What our figure buys instead is engine breadth and a rerun cadence, and no arithmetic converts one into the other, which is exactly the point: 100 prompts on 3 engines and 25 prompts on 5 engines twice a week are different products, and a single sorted column hides that rather than revealing it.

The unit problem gets worse once you ask what a prompt result means. We published a count on 29 August 2026 from this repository's own engine registry showing that 3 of the 8 engines we ask cannot search the web at all: five are recorded as retrieving live and returning the sources they grounded in, three are recorded as answering from prior knowledge with no web access and no sources, and a free run asks exactly one engine, which is one of the three. A mention produced by an engine with no retrieval is a fact about training data from months ago. A mention produced by a retrieving engine is a fact about what a search index surfaced this week. Averaging them into one visibility percentage, which the shape of every dashboard in this category invites, blends the two into a number that means neither. The prompts this product runs and what each one is for are on the prompt set this product runs.

Crawler access tools: your logs, your edge, or any public URL

The other half of the category answers whether a bot can get the page, and it divides again by where the evidence comes from. Three sources exist and they see different things.

Your own server logs are the strongest evidence available for whether a named crawler actually fetched a page, because they record requests that really happened rather than requests a tester made. Screaming Frog's SEO Log File Analyser is the established desktop tool here, and its product page, read on 1 September 2026, states a free version limited to a log event limit of 1,000 lines and a project limit of 1, with the paid licence at 139 US dollars per year and the same page also quoting 99 pounds per year. The page advertises automatic verification of search engine bots and the ability to view IPs spoofing requests, which is the check that separates a request from a crawler from a request from something wearing its name. That capability is straightforwardly better evidence than anything an external scanner can produce, including this one, and the page is at screamingfrog.co.uk/log-file-analyser if you want to read it yourself. No external tool can see your logs, so no external tool can tell you what actually visited.

The edge is the second source, and it is where the price of this capability collapsed. Cloudflare's AI Crawl Control documentation, last updated 14 August 2026, describes a product to monitor and control how AI services access your website content, showing which AI services access it, allowing allow and block rules per crawler, and tracking robots.txt compliance including which crawlers violate the directives. The documentation states it is available on all plans. For any site already behind Cloudflare, the per crawler request picture that several paid products put in a dashboard is therefore included at no additional cost, which is worth knowing before paying for it twice. Cloudflare is also changing what those defaults are, which we reported separately when Cloudflare's new default depends on whether the page carries ads landed.

The third source is a fetch of a public URL, which is what Lantad does and what every free checker does. It is the weakest source in one specific respect, that it observes a request the tool made rather than one a crawler made, and the strongest in another, that anybody can run it against any site without credentials, permission or a login, and get the same answer twice. That is the reproducibility criterion from the top of this post, and it is the only column in the hero table where the crawler side of this product wins outright. You can see the individual pieces of it without buying anything, in what GPTBot sees on your page, in the robots.txt tester, and in the crawler token list that records which tokens exist and which vendor owns each.

Where Lantad fits, and the three things it cannot tell you

This post is published by one of the products in the table, so the useful thing to do here is state the limits rather than the pitch.

Lantad grades what a public URL returns, and the composite is four weighted parts recorded in SCORE_WEIGHTS in core/src/config.ts: prose parity at 0.5, access at 0.25, structure at 0.15 and schema at 0.1. Those weights are a decision about what to count, not a measured discovery that parity is worth half your visibility, and nobody should read them as evidence about how answer engines behave. The deeper reading of structured data that ships alongside the grade, Entity Confidence, is deliberately worth zero of those hundred points, because it is a diagnostic rather than a component of the score. The aggregate picture we are willing to publish from real scans is on the research page.

Three things this product cannot tell you, all of which matter when choosing between the halves of this category. It cannot see your server logs, so it cannot tell you whether GPTBot actually fetched a page, only whether it would be served one now: the log analyser named above genuinely beats it there and a serious diagnosis uses both. It cannot see settings held inside your own accounts, such as a Search Console property level exclusion, because those change nothing about the bytes your server returns. And on the answer tracking half it sells a smaller panel than the dedicated trackers do, 25 prompts per run at the top published tier against 100 tracked at Profound's equivalently priced plan, so an agency whose deliverable is a wide monthly prompt panel across many brands is buying the wrong tool here and should buy one of the trackers.

The buying rule that falls out of all this is duller than any ranking. Work out which of the two questions you have. If nobody has ever confirmed that a crawler can read your pages, start on the crawler side, because a tracker will otherwise sell you a year of correct reports saying nothing mentions you. If your pages are demonstrably readable and the open question is which brands the models name, buy a tracker and buy the widest panel you can justify, since panel breadth is the thing that decides whether the number moves for a real reason or for sampling noise. If you already sit behind a CDN, look at what it gives you free before paying anyone for the same view. And whichever half you land in, the platform specific mechanics of getting quoted are a separate body of work from either measurement, which is what how to get cited collects.