How To Run A DIY AI Visibility Check In One Afternoon

A free AI visibility checker hands you a score, not a baseline. In one afternoon you can build a better one by hand. Write twenty buyer questions, run each across ChatGPT, Perplexity and Google AI Overviews, and score mentions, citations and competitors in a sheet you can rerun next month.

This holds for a B2B SaaS team that wants a first baseline this week and has no audit budget yet.

One pass of any prompt set is noisy, so read the result as a starting point and move to a paid audit once the triggers in the decision rule apply.

What An AI Visibility Check Measures

A checker gives you a single number. That number hides four separate questions, and each one needs its own column in your sheet.

Here is what each of the four measures records.

• Mention: The brand name appears anywhere in the answer, yes or no.
• Citation: The answer links to a page on your own domain as a source, yes or no.
• Position: Where the brand sits in the answer. First recommendation, one of several, or an aside.
• Competitor presence: Which rival brands the same answer names.

A search ranking is one page in one slot. An AI answer is a paragraph that may name three brands, cite five sources and change tomorrow. That is why a position-one screenshot tells you very little here, and why the same SEO habits only partly carry over.

A manual check is simply the work of asking the questions buyers ask in each answer engine, noting whether your brand appears and what gets cited, and repeating it monthly or quarterly. Nothing in it needs a licence. It needs a fixed list of questions and the discipline to score every answer the same way.

A single checker score is not a baseline because you cannot see what was asked, how many times, or why the number moved. A baseline is something you can rerun with identical inputs and compare. The rest of this guide builds that.

The Afternoon Plan At A Glance

Read this table first. If the time boxes look workable, the rest of the article is the detail behind each row.

The five blocks add up to 220 minutes, about three hours and forty minutes. At the end you hold a dated sheet of sixty scored answers, a tally of the sources the engines trust, and a first baseline you can rerun next month.

Step 1: Build A Prompt Panel Of Real Buyer Questions

The panel is the whole check. A weak panel gives you a confident wrong answer, so spend the first forty-five minutes here and do not rush it.

Write twenty prompts, four for each of five question types. Twenty is the right size for one afternoon because it gives sixty answers across three engines, which is about what one person can run and score without cutting corners. A larger panel adds coverage but also hours. Add to it later, at a quarter boundary, once the first sheet has told you where the gaps are.

Follow these steps.

1. List the five question types on a blank sheet and leave four rows under each.
2. Write each prompt the way a buyer would type it, in plain words, not in marketing language.
3. Keep one intent per prompt. A prompt that asks for a category, a price and a comparison at once cannot be scored cleanly.
4. Check every unbranded prompt for your own brand name and remove it.
5. Read the twenty prompts aloud and cut any that a real buyer would never say.

Unbranded prompts carry no brand name. That single rule matters more than any other, because a prompt that names your brand tests whether the engine knows you exist, not whether it would recommend you to a stranger.

The template below gives one worked prompt per type. Swap in your own category and competitors.

Four brand-direct prompts is the only place your name goes in. Those prompts check accuracy rather than reach, so a wrong or outdated description of your product shows up here first.

Wording changes results more than most people expect. If you want to see how far, read our notes on prompt variation testing, then lock your wording and do not touch it between runs.

Step 2: Run The Panel Across The Main Answer Engines

Run ChatGPT, Perplexity and Google AI Overviews first. These are the three core engines for this method, and running the same panel through all three lets you compare them side by side. Gemini and Copilot are optional extras if you have time left.

Hand-running beats guessing, but it does not scale. Sixty answers is fine for one afternoon. Six hundred is a job for a tool, and Step 5 and the decision rule explain where that line sits.

Use this run protocol for every prompt.

1. Open a fresh session for each prompt so earlier answers cannot colour the next one.
2. Turn web search on, because an answer without search shows the model's memory, not what it cites.
3. Use the same location for every run and stay logged out where the engine allows it.
4. Paste the exact prompt from the sheet without edits.
5. Copy the full answer into the sheet, not a summary of it.
6. Copy every cited link, including the ones hidden behind footnote marks.
7. Stamp the date and the engine beside the answer.
8. Move to the next prompt only after the row is complete.

ChatGPT goes first because most teams already know where their brand stands there, or think they do. Our guide to tracking brand visibility in ChatGPT covers the engine in more depth.

Google AI Overviews are the answer that sits above the normal results. If getting cited in those answers is the goal, our Answer Engine Optimization service is built around it.

Expect the run to feel repetitive by prompt twelve. That is normal. The temptation is to skim the answer and tick a box, which is exactly how mentions get missed. Slow down on the answers where your brand does not appear, because those are the ones the sheet exists to explain.

Step 3: Cross-Check With A Free AI Visibility Checker

Run a free checker after the manual pass, not before. Used first, it anchors you on its score. Used second, it is a quick sanity check on your own sheet.

Free options exist and are worth a few minutes. SEO Site Checkup's free LLM visibility checker covers six AI platforms in one scan, needs no signup and returns a report in under 30 seconds. Ahrefs offers a free AI Visibility Checker too. Run one or two, then compare what they report with what your sheet shows. This is a cross-check, not a ranking of tools, so there is no list here to rank.

Use a checker as a cross-check after the manual pass, and treat any disagreement as a prompt to look at the specific answers in your sheet.

The right column is why the manual sheet exists. A checker can tell you roughly where you stand. It cannot tell you why, and it cannot tell you whether the result will hold next week.

Step 4: Score Mentions, Citations, And Competitors

Scoring turns sixty pasted answers into numbers you can compare month to month. Build the sheet with one row per prompt and engine pair, and use the columns below exactly as written.

Score each row the same way every time. A mention counts only if the brand name is in the answer text, not if it only appears in a cited page title. Cite a domain only when a link in the answer points to it.

Here is the worked calculation. It is an illustration with made-up counts, so replace them with your own.

• Panel size: 20 prompts across 3 engines gives 60 answers.
• Mentions: The brand is named in 15 of the 60 answers, which is a 25 percent mention rate.
• Citations: The brand's own domain is cited in 9 of the 60 answers, which is a 15 percent cited rate.
• Competitor share: Rival A is named in 33 of the 60 answers, which is a 55 percent competitor share.

In this example, the brand is named in a quarter of answers, cited in about one in seven, and sits well behind Rival A. The gap between the 25 percent mention rate and the 15 percent cited rate is also useful. It says the engines know the name but rarely send readers to the site.

Calculate cited rate and competitor share for every competitor you listed. Divide the count by 60, not by the number of answers that mentioned anyone.

So what is a good score? No universal benchmark exists, and anyone quoting one is selling something. The useful comparisons are your own mention rate month on month and your named competitors in the same sheet. Beat your own last month and close the gap to the rival you chose.

If you want a longer scoring template, the step-by-step LLM audit playbook goes deeper, and AI search visibility metrics and KPIs explains how each measure is defined.

Step 5: Read The Sources Behind The Citations

Mentions tell you whether you appear. Sources tell you why. Take the source domains column from the sheet and tally them by type for each engine.

A source gap is a source the engines cite for your competitors that does not mention you. For example, a competitor named in four answers that all cite the same third-party roundup, where your brand is absent from the roundup.

Follow these steps.

1. Copy the source domains for one engine into a separate tab.
2. Sort each domain into one of four types. Own site, competitor sites, review or comparison sites, and forums or news.
3. Count how many times each type appears.
4. Mark every competitor mention that traces back to a third-party page you are not on.
5. Repeat for the other two engines and compare the patterns.

Here is what a small tally might look like. The counts are made up to show how to read it.

The review row is the one to chase. If the engines lean on a roundup that omits you, that is a gap your own pages do not close, so check whether you can be listed there accurately.

This is also where a hand check hits its limit. A tally by domain type is easy. Knowing which exact page on which exact site drove which answer, across hundreds of prompts, is URL-level attribution, and it is where tools start. Our look at page-level citation attribution tools covers that point.

Rule Out A Blocked Crawler First

Run this check before you blame your content. A site that blocks AI crawlers can be absent from AI answers for that reason alone, so crawler access comes before any judgement about quality.

Work through these three items.

• Read your robots.txt file and confirm no rule blocks the crawlers you want reading the site.
• Check your host or CDN bot rules, because a firewall setting can block a crawler that robots.txt allows.
• Confirm the pages you want cited do not carry a noindex tag.

If all three are clean, the gap is likely about content or sources, and the earlier steps point you to which. For the wider picture of how models read a site, see our LLM SEO service page.

Common Mistakes That Skew A DIY Check

A hand check misleads in predictable ways. These are the ones to avoid.

• Brand name in the prompt: It turns a reach test into a recognition test. Remove the name from every prompt except the brand-direct ones.
• Logged in with memory on: Personalised context shapes the answer. Run logged out, or with memory switched off, and use the same setup every time.
• Counting any mention as a win: A mention as a cautionary example is not a recommendation. Score position and accuracy, not only presence.
• Trusting one run: A single pass is a first reading. Repeat a few prompts to see the spread before you act.
• Ignoring a wrong description: An engine that names you but gets your product wrong is costing you. Flag it in the accuracy column and fix the source it relies on.
• Changing the wording between months: New wording is a new test. Keep the panel fixed and add prompts only at a quarter boundary.
• Forgetting model updates and personalised answers: Both can move your results without any change to your site, so the sheet needs a date on every row and a note when an engine visibly changes.

Why One Run Is Noise And How Many Runs You Need

Run the same prompt twice and the answers differ, because answers are sampled. Differences within one prompt are mostly noise. Differences across different prompts are the signal you are after.

The size of that noise has been measured. Evertune's study of 10,700 ChatGPT prompts puts the margin of error for one brand at roughly 27 percentage points at 5 runs, 12 at 12 runs and 6 at 100 runs, with diminishing returns beyond that. The table sets those figures side by side.

Repeat runs are not the only way to tighten a reading. Meikai's ten-day test of 500 prompts used 500 fixed US English prompts across five engines, 250,000 responses in all, from 13 to 22 July 2026. It found that one daily run per prompt and platform pair gave the same pooled ten-day estimate as ten daily runs. Extra runs narrowed the error range without moving the pooled estimate.

The wider point has an academic anchor. An April 2026 arXiv paper argues that measuring AI search visibility once is unreliable, and that visibility should be measured repeatedly and treated as a distribution.

For the afternoon, the practical rule is this. Treat the first full pass of the panel as a first reading, repeat five spot-check prompts a few extra times to see how wide the spread is, and then rerun the whole panel on different days. Tools that log answer changes automate exactly this, and our list of answer change detection tools is the place to start.

Set A Monthly Cadence So The Baseline Holds

A baseline is only worth having if it repeats. Turn the afternoon into a routine with these steps.

1. Reuse the same panel, the same wording and the same engines every time.
2. Run it on the same day each month, so the gap between readings is even.
3. Put each run in a new tab and leave the first sheet untouched.
4. Compare the new tab with the previous one, prompt by prompt, before you look at the totals.

Run monthly with the same panel. The single exception is a quarter boundary, when you may add new prompts, as long as you keep the old set so the trend line survives.

Hand reruns have a cost. When copying answers into a sheet every month starts to cost more than it saves, Semrush tracks AI visibility on a schedule, which covers the repeat-run part of this job without the manual copying.

If you move to a tool, test it before you pay. Our AI visibility platform buyer guide recommends a 14-day proof-of-value test using 25 to 50 prompts, which your twenty-prompt panel is a good start towards.

When A DIY Check Is Enough and When To Book A Paid Audit

The hand check is not a lesser audit. It answers a narrower question, and for some decisions that is all you need. The table lays out where the line falls.

These triggers are our own house rule, not a market statistic, and they carry no invented thresholds. Each one is something you can see in the sheet.

Book a paid audit when any one of these five triggers fires:

1. Your mention rate is close to zero while the sheet shows named competitors appearing in the same answers.
2. Repeat runs of the panel disagree with each other across many prompts, not just one or two.
3. You need to know which exact URLs the engines cite, at a volume the sheet cannot hold.
4. The competitor set you need to track is wider than the sheet can hold, or you are covering several brands or regions.
5. Leadership wants monthly reporting rather than a one-off reading.

If none of the five fires, the sheet is probably enough for now. Repeat it, fix what it shows, and look again next month.

If one or more fires, read the full AI visibility audit guide for what an audit covers. When the trigger is a competitor set the sheet cannot hold, competitor benchmarking tools show what a wider comparison looks like.