How to Evaluate GEO Tools Without a Ranking List

Most "best GEO tools" lists rank vendors by features, price or affiliate payout. None of those tells you whether a feature does anything. The question that matters is simpler: does each thing the tool measures or recommends have evidence behind it?

The evidence is narrower than the market suggests. A July 2026 critical survey of 45 GEO studies (arXiv:2607.14035) concluded that already-retrieved content can causally alter its citation or use, but no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability. Most tools sell the second thing.

So you need a rubric, not a list. This page gives you one you can apply to any tool in an afternoon, without trusting anyone's ranking, including ours. If you're still deciding what job the tool should do, start with how to choose an AI visibility tracking tool.

An AI answer passes through four stages. Knowing them tells you which features are measurable and which are guesses.

• Parametric: what the model knows with no web access, frozen at a training cutoff. Nobody changes this quickly.
• Access: whether a crawler may fetch your page at all. You and your CDN control this.
• Retrieval: whether your page enters the model's context for a given query. The engine controls this, partly.
• Generation: which retrieved pages the model uses and cites. The model controls this.

A tool can check access directly. It can request your pages as each AI crawler, read your robots.txt per bot, and diff raw HTML against rendered HTML. Vercel's crawler study found that ChatGPT and Claude crawlers fetch JavaScript files but don't execute them, so content that appears only after JavaScript runs may be invisible to them.

A tool cannot see retrieval from outside. It can't observe the real queries an engine runs, because OpenAI rewrites prompts into its own search queries. It can't know your position inside an answer either. As of October 2026, Google's Search Console report for AI Overviews and AI Mode shows impressions only, with no position data. Any tool that claims to see these things is estimating, and should say so.

Score each tool from 0 to 2 on every test. A tool scoring 0 on test 5 is out, regardless of its total.

A tool scoring 8 or more is worth a trial. Below 6, its numbers will be hard to act on. The sections below explain what a pass looks like for each test.

Go through the tool's recommendations one by one and sort them into three buckets.

Strong evidence

These are access fixes, and they work by removing blocks:

• Server-side rendering of the facts that matter (prices, product names, definitions).
• Per-bot robots.txt permission. OpenAI's crawler docs state each setting is independent: GPTBot governs training, OAI-SearchBot governs ChatGPT search.
• Network reachability. A firewall or CDN can return a 403 to a crawler your robots.txt allows.
• Index and snippet eligibility. Google states a page must be indexed and eligible for a snippet to appear as a supporting link in AI Overviews or AI Mode, with no additional requirements.
• Accidental suppression through noindex, nosnippet, max-snippet:0 or data-nosnippet.

A tool that checks these earns full marks on test 4.

Weak or contested evidence

Freshness, IndexNow pings and Bing index presence have a documented mechanism or an assertion, but no measured effect on AI citation. A good tool labels them as hygiene.

Evidence against

Schema as a citation lever, llms.txt, chunking, "AI-friendly" phrasing and E-E-A-T scores. Ahrefs' matched-control study of 1,885 pages found AI Overview citations fell 4.6% after adding schema. Google's optimization guide says Search ignores llms.txt and rejects writing in a special style for AI.

If most of a tool's recommendations land in the third bucket, you'll spend your roadmap on tasks that don't move anything. Our guide to retiring unsupported GEO tasks covers what to cut.

Engines behave differently, so a single number misleads. Google says AI Overviews and AI Mode may use different models and techniques, and they return different links. Citation units don't match either: Anthropic's API cites text spans, while Perplexity cites whole sources by index. There is no shared denominator to average.

Retrieval also varies per answer. As of October 2026, ChatGPT searches at the model's discretion, Claude.ai search is a user toggle, and Perplexity's Agent API needs web search switched on. An answer without search reflects training data only, which no on-site change reaches this quarter.

What to demand:

• Results for each engine on its own line.
• The raw answer text, so you can check what was said.
• A flag per answer showing whether search ran, plus the sources it used.
• Where possible, retrieved sources listed apart from cited ones. OpenAI's API exposes both, and the retrieved set is usually larger.

Be fair to vendors here. Any single AI visibility score, including the one 99Visibility publishes, averages answers given with and without live retrieval. The pass condition is disclosure, not a perfect metric. A tool that explains what its score is made of passes. A tool that presents its score as a measurement of AI visibility does not.

Test 5 is a filter. Read the tool's homepage, pricing page and sample reports. Any of these claims is a fail:

• "Up to 40% more visibility." That figure comes from the original GEO paper, measured on an engine the authors built with sources already loaded into context.
• "Your AI rank" or a position in Google AI. No position data exists to check it against.
• "Improve your E-E-A-T score." Google says E-E-A-T isn't a specific ranking factor. There is no score.
• "Build domain authority for AI." Domain Authority is a Moz metric. No engine computes it.
• "Add FAQ schema to get cited." FAQ rich results were deprecated in May 2026, and the matched test above found a decline.
• "Allow GPTBot so ChatGPT finds you." GPTBot governs training only.
• "Ping IndexNow and AI cites you faster." The spec never mentions AI. A 200 response confirms receipt only.
• Guaranteed citations or mentions. No tool controls what a model says.

Case studies deserve the same scrutiny. Read whether to trust GEO case studies that show big citation gains before a before-and-after chart sways you.

Work in this order. Each step removes tools before you spend time on them.

1. Decide the job. Tracking tools report what engines say. Fixing tools find what blocks you. Read tracking tools vs fixing tools and pick one, or both.
2. Run the claims filter. Spend ten minutes per vendor site. Drop any tool that fails test 5.
3. Run a free trial or audit. Prefer tools that let you see output before a sales call. This list of tools with a free first audit and no demo helps.
4. Use the same three buyer questions everywhere. Pick questions a real customer would ask, such as "best accounting software for a five-person firm". Run them through each tool. Compare raw answers, not scores.
5. Check one access result by hand. Pick a tool's crawler finding and verify it. You can request a page as each AI crawler with the free AI crawler checker and compare the result with the tool's report.
6. Score the rubric, then look at price. Cost only matters among tools that pass. See what AI visibility tools cost per month for typical ranges.

Say you run a hypothetical dental clinic in Austin. Step 4 might show one tool reporting a strong score while its raw ChatGPT answers never name you. That gap is the finding.

• Comparing scores across tools. Each vendor builds its score differently. A 70 in one tool and a 40 in another may describe the same answers.
• Trusting prompt volume. Engines rewrite your prompt into their own queries, so a tracked prompt is a proxy. More prompts give a bigger sample, not a truer one.
• Ignoring the score formula. Ask whether the accuracy or visibility score divides by the number of answers. Established factuality metrics such as FActScore normalise by a count. A score that falls simply because you tracked more prompts is broken.
• Skipping access checks. A clean robots.txt means little if your CDN blocks the crawler. Cloudflare changed its default AI crawler settings on 2026-09-15 for new domains and free-tier customers on pages showing ads.
• Buying recommendations you'd reject on evidence. If the tool's top suggestion is schema or llms.txt, its prioritisation is off.
• Expecting a tool to control outcomes. A good tool tells you what's broken and what fixing it involves. What the models then say is outside anyone's control.