Disclosure: This evaluation guide is published by SkillAEO, which offers its own free anonymous AI visibility check and paid audit plans. We have not inspected proprietary backend code for Searchable's AEO Grader or other third-party graders; any unverified internal weighting is treated as Unknown rather than guessed.
If you just ran Searchable AEO Grader, HubSpot AI Search Grader, or another free AI visibility score tool and wondered, "How accurate is this number?"—the honest answer comes down to separating single-run observation fidelity from market-wide prediction.
A free AEO grader can accurately report whether a specific model mentioned your brand on a handful of prompts at one moment in time. No free grader can measure your true share of real user conversations, because AI assistants personalize context, vary answers across identical runs, and do not publish domain-level impression logs the way Google Search Console does.
Verify with inspectable raw answers
Compare any grader score against raw prompt evidence
Run a 5-question GPT check on the same public domain and inspect the exact generated prompts, complete answers, cited URLs, and separate HTML readiness checks.
Anonymous scope and boundaries
Without an account, you can run up to 3 anonymous audits per day, 1 per domain, subject to daily capacity. Each anonymous audit checks 5 generated questions with 1 AI engine: GPT, and does not include a full report. Results vary by model, search mode, locale, prompt, and time.
Two very different meanings of "accuracy" in an AEO grader
When marketers ask whether an AEO grader is accurate, they usually mix together two separate questions:
| Accuracy question | What can be 100% verified | Where free graders mislead buyers |
|---|---|---|
| 1. Did the tool honestly record what the model output? (Observational accuracy) | Whether the report shows the exact prompt text, timestamp, AI family (GPT, Claude, Gemini, Perplexity), complete response text, and clickable cited URLs | Hiding raw prompts and answers behind an opaque 42/100 badge or email gate so you cannot tell whether a "mention" was positive, outdated, or about a similarly named company |
| 2. Does that score represent what real buyers see every day? (Statistical & predictive accuracy) | Almost never from a single free run—a 5-to-20 prompt snapshot is a narrow diagnostic sample, not a representative buyer census | Presenting a one-time sample score as a universal "AI market share" or promising that fixing checklist items will raise citations by a fixed percentage |
7 checks to audit Searchable AEO Grader (or any AI visibility checker)
Before you bring a grader screenshot into a leadership meeting or buy a monthly subscription based on a low grade, run the report through these seven checks:
1. Can you read the exact prompts word for word?
Prompt phrasing dominates whether a brand appears. A query that includes your brand name ("What does Acme do?") tests entity recognition, whereas a category query (" project reporting tools for small retail teams") tests competitive recall. If a grader blends branded and unbranded prompts without showing the wording, the headline score cannot be interpreted safely.
2. What is the exact denominator?
A 20% visibility score on 5 prompts means 1 prompt mentioned your brand. On a small denominator, a single stochastic variation moves the score by 20 percentage points. Always record mentions / total valid runs rather than quoting percentages in isolation.
3. Did the grader use live web search or ungrounded model memory?
An ungrounded LLM completion answers solely from pre-cutoff training weights, while web-grounded search modes (ChatGPT Search, Perplexity, Google AI Overviews) retrieve live web pages and attach source URLs. A newly launched startup often scores zero in ungrounded memory while appearing accurately in live search retrieval—or vice versa if robots.txt blocks search crawlers.
4. Can you inspect raw answers and cited URLs?
Entity collisions are common. If your company shares a name with an open-source library, local agency, or dictionary noun, regex-based brand counters can credit you for an unrelated mention—or miss your product when the model cites your domain without capitalizing your brand name. Always inspect the full paragraph and cited URLs (see how our public sample report separates mentions, citations, and accuracy).
5. Are observed AI answers kept separate from static HTML checks?
Some graders combine HTML checklist heuristics (meta tag length, heading counts, Schema.org presence, or optional files like /llms.txt) with LLM answer checks into one blended grade. That creates two distortions:
- A well-known brand with messy HTML can look "failing" on the grader even when ChatGPT and Perplexity cite it constantly.
- Worse, some graders penalize sites for missing
/llms.txtor custom agent manifests even though Google's AI search documentation explicitly states Google Search does not usellms.txt(and at SkillAEO, optionalllms.txtand experimentalagent.jsonchecks carry zero weight in readiness scoring; see our scoring methodology).
6. What happens when you test the exact same domain twice?
Run the same domain on two consecutive days—or compare the grader's output against a manual check in the ChatGPT visibility check. Moderate run-to-run variation is normal in generative models; if a vendor claims its single-run grade is deterministic, treat that claim skeptically.
7. Are recommendations tied to verifiable primary documentation?
Healthy audit output points to concrete, low-regret improvements: clarifying homepage category copy, unblocking OAI-SearchBot / PerplexityBot / Googlebot in robots.txt (testable via our AI crawler & robots.txt checker), publishing transparent pricing and comparison pages, and keeping Schema.org JSON-LD aligned with visible HTML. Reject fear-based pitches claiming your site is "invisible to AI" unless you buy proprietary files.
How free AEO graders compare on evidence transparency
| Check dimension | Opaque lead-gen grader pattern | HubSpot AI Search Grader (public page checked 2026-08-18) | SkillAEO anonymous check |
|---|---|---|---|
| Account required | Often email-gated before details | Public page says free and no account required | No account required (up to 3/day, 1/domain, subject to daily capacity) |
| Prompt & answer visibility | Summary score only | Public page describes brand-dimension scores; complete raw answers not verified in our public-page check | Shows the 5 generated GPT questions, raw responses, and cited URLs |
| Technical readiness separation | Often blended into one score or omitted | Brand-focused snapshot | Reported separately from observed AI visibility; optional llms.txt carries 0 score weight |
| Best use | Quick curiosity check | Lightweight multi-surface brand snapshot | Inspecting concrete prompt answers alongside verifiable public-page checks |
For a dedicated comparison with HubSpot's free tool, read Skillaeo vs HubSpot AEO Grader, or browse all 12 reviewed products in Best AEO Tools.
Frequently Asked Questions
Why did two free AEO graders give my domain completely different scores?
Different graders ask different generated prompts, query different AI models or search modes, weight brand mentions versus link citations differently, and may mix static HTML rules into the score. Compare the underlying prompt text and raw responses rather than the top-line numbers.
Is a low AEO grader score proof that nobody finds us in ChatGPT?
No. A low score on 5 or 10 generic category prompts only shows that your brand was absent in that specific sample. Check your own web analytics and first-touch referrer logs (such as chatgpt.com and perplexity.ai referrals) alongside manual prompt checks for your core buyer use cases.
Should I fix every technical warning a free grader flags?
Verify each warning against official search and crawler documentation first. Fix real crawl blocks (robots.txt, broken canonicals, missing sitemap.xml, vague homepage copy), and treat optional proposals like /llms.txt as editorial conveniences rather than mandatory ranking signals.
How does SkillAEO keep its anonymous audit verifiable?
SkillAEO's anonymous audit uses GPT for 5 questions without an account (up to 3 anonymous audits per day and 1 per domain, subject to daily capacity, and does not include a full report). It exposes the sampled questions, answers, and cited URLs, reports public-page readiness separately, and documents exact weights on the methodology page.
