How to choose an AI visibility tool.

Four different jobs, and most products now do more than one. A tool can track answers: run your buyer questions through AI engines and record who was mentioned, cited and recommended. It can diagnose them, showing which sources an answer leaned on and why you’re missing. It can help produce the fix, with a queue of work, drafts and a path to publishing. Or it can check retrievability: whether engines can fetch and read your pages at all. Name the job before you compare features.

Most products in this category now go past the chart. They recommend actions, draft content, connect to a CMS or follow up on what you shipped, and each packages that differently. Our comparison pages document six of them from their own product pages. So a feature grid is a weak guide: two products with the same checkboxes can leave very different amounts of the work with your team.

Whatever the job, it rests on a measurement, and the measurement is where tools differ most quietly. An answer can cite your page and recommend your competitor, so a tool that blends mentions, citations and recommendations into one score hides the row you’d act on. Look for tools that keep them apart.

Maybe not, and a few checks will tell you. Start with a free first look: ask ChatGPT what it tells buyers about you, open every source it cites, and sort each claim into supported, unsupported, stale or judgment. Then fetch your most important pages with JavaScript switched off. The major AI crawlers in Vercel and MERJ’s study fetched JavaScript files without running them, so copy that appears only after scripts run never reaches them, and no subscription fixes that.

Next, build a baseline by hand. Pick five to ten questions buyers actually ask, write each two or three ways, run them on two engines once a week for a few weeks, and log who was mentioned, cited and recommended, with every source. A spreadsheet protocol is enough. Our own rule is breadth over repeats, and that’s a habit rather than a published standard. The baseline shows which job you actually need done, and gives you something to hold every trial against.

Hold off on buying if you have no pages that answer buyer questions yet, if nobody can act on a finding this quarter, or if your market comes down to a few questions you can check yourself. Buy when the hand check breaks: when the question set outgrows the spreadsheet, when you need an archive of every answer, or when logging crowds out acting on what you find. Either way, freeze the questions and book the rerun before you start a trial, so every tool is judged against the same baseline.

Ask for the sample under every number, then check it against your own. AI answers change from run to run: in a 2026 variance-components study, one answer from one model in one phrasing ranked brands with a reliability of about 0.01, and precision came from spreading measurement across engines and phrasings more than from repeating one prompt. A tool’s number is only as good as the questions, runs, engines and dates behind it. In a trial, ask to see:

• How many questions, runs per question, engines and dates produced each number, with the sample size shown.
• Results per engine and per question, with their spread, before any blended score.
• The full answer text and every cited source, not just a flag that you were mentioned.
• A raw export you can open yourself, and what happens to your history if you cancel.
• Which model versions produced the last quarter of data.

Then run one of your baseline questions by hand next to the trial. If the tool’s history looks smoother than your own runs, ask how it was smoothed. Hold published numbers to the same standard: the IAB’s framework treats fewer than 50 queries as exploratory and asks providers to show results per platform with their variance, and a vendor benchmark is a hypothesis until it discloses its brands, prompts, engines and runs.

It depends on who will act on what it finds. A tracker earns its keep when you already have a team that ships: it flags drops faster than hand checks, benchmarks a fixed set of competitors and produces the report, and your people do the rest. If nobody turns a finding into a change, a better chart leaves the gap where it is.

If you want the tool to help with the work, judge it on the handoffs, not the demo. A gap becomes a brief only when a page you own could fill it, so check that each recommendation traces back to the answers and sources behind it. Then check the seams: are full answers kept, can you see why an action was chosen, does an approval record who said yes and why, and does the tool rerun the question after you publish? Any gap is a stage you build yourself.

Whichever you buy, keep a person on anything a buyer will read. Automate the watching and the internal drafts, and let a named reviewer own the claims and the publish button. An approval that nobody can refuse isn’t a control.

On one of your own questions, from each vendor’s own documentation. Feature grids flatten the differences that matter, and roundups age fast, because this category repackages faster than any article can track, ours included. Put the same buyer question through every tool on your shortlist and inspect the original answers, sources and dates each one keeps.

Then compare on what a pricing page won’t show: how deeply each question is sampled, what each snapshot keeps, which engines and model versions it reaches, and what seats and exports really cost. Price the configuration you actually need, because 50 questions asked once is a different product from 50 questions asked in several phrasings, on every engine you need, week after week.

Add one hard case to the trial: a competitor that’s genuinely better for some buyers, a stale source, or brand copy that contradicts your docs, and see how each tool handles it. When you read comparisons, ours included, check the claims against each vendor’s current pages. Our comparison pages name their sources and the date each was reviewed, for Profound, Peec AI, Scrunch, Goodie, AirOps and Jasper.

Make an engine cite you. Engines decide what to cite and recommend, and a tool that promises otherwise is selling against the evidence. Much of what’s sold as AI optimization is thinner than the pitch, and most of what moves answers happens off the dashboard: whether engines can read your pages, whether a page carries evidence worth quoting, and whether sources you don’t own vouch for you.

The evidence on tactics is thin. When C-SEO Bench retested ten conversational-SEO methods across two tasks and six domains, most proved largely ineffective and often harmful to citation rank, and gains shrank as more sites adopted them. The strongest case left is for evidence: in the original GEO study, the best methods added quotations, statistics and sources, and improved Position-Adjusted Word Count by up to 41% in the authors’ own benchmark, not in live engines. Treat added evidence as a sound default whose citation lift is still unproven, and ask any tool that scores your writing what the score is based on and whether it has been tested against real citations.

And much of any answer sits beyond your own pages. In Profound’s analysis of 11.84 billion citations, about 57% pointed to company websites, yours and your competitors’, and the rest to editorial coverage and platforms like Reddit and YouTube. No tool publishes you into a review site or a community thread. Earning those takes reviews, partners and PR, and the most a tool can do is show you where they matter.

Buy the measurement unless measuring AI answers is close to your core product, and keep the judgment in-house either way. A credible build is a data pipeline, not a script: questions run on a schedule, statistics that separate change from noise, an archive of every answer, and upkeep every time an engine or crawler changes. The dashboard is the cheap half.

Building makes sense when measurement is your product, when your questions are too unusual for a vendor’s setup, when you already run experiments and think in confidence intervals, or when policy forbids sending your competitive questions to a third party. There’s a middle path too: buy the measurement, and build thin tools on its exports.

Waiting is also a choice, and sometimes the right one. On a small budget, run the fetch test and a hand baseline first, then fund one narrow territory you can answer better than anyone before you buy a tool. Whatever you buy, insist on a raw export and own everything you publish, so switching never resets your baseline to zero.