Reading your Results

Reading a baseline means cutting the results three ways, by engine, by prompt intent and by cited source, and turning each real pattern into a written finding with evidence, a likely cause and an action. A baseline full of percentages doesn't change anything by itself. Three good findings do. Lesson 7 of GEO 101 shows how to make each cut, how to deal with accuracy errors, how to tell a pattern from noise, and what MentionX data says about citation sources that should change how you read your own.

In this lesson

• Before you start: check the data
• Cut 1: by engine
• Cut 2: by intent
• Cut 3: by source
• From MentionX data: citations are a long tail
• Putting the cuts together
• Accuracy errors
• Signal or noise?
• Writing findings
• Worked example: Ledgerline's three findings
• Mistakes when reading results
• Prioritising and presenting
• Exercise
• Quiz
• Sources

Anything marked "Our view" is our opinion. Ledgerline is the fictional company used throughout this course.

Before you start: check the data

Spend ten minutes on quality before you analyse anything:

• Spot-check 20 answers against your scoring rules. If a tool did the scoring, look for missed name variants and mentions that should have been recommendations.
• Look for empty or refused answers and count them. They belong in the denominator, but a cluster of them on one engine is itself a finding.
• Confirm every row has engine, mode, prompt ID and date. Rows you can't place can't be used.
• Check competitor matching. If one competitor's numbers look strangely low, a name variant is usually missing.

Cut 1: by engine

Lay out your four metrics per engine, with n, next to your top competitor.

For ChatGPT, OpenAI documents that sites which opt out of OAI-SearchBot aren't shown in ChatGPT search answers (OpenAI crawlers). For Google's AI features, a page has to be indexed and eligible for a snippet to be a supporting link (Google Search Central). Rule those out before you draw any content conclusions.

These are common interpretations, not rules. Check each one against the actual answers.

Cut 2: by intent

Next, break the rates down by prompt intent.

The intent cut usually gives the most actionable finding. "We lose problem prompts" points to a specific type of page. "Our mention rate is 28%" points to nothing.

Look at competitor rates by intent too. The intent where the gap to the leader is biggest is usually the first one worth working on.

Cut 3: by source

List every cited domain on your commercial prompts (best-of, comparison, alternative), with:

• How many answers cited it
• Whether that page mentions you, mentions you accurately, or doesn't mention you
• Whether it mentions your leading competitor

Then sort the domains into the four types from Lesson 3: own, about you, competitor, neutral. The domains cited often that don't mention you are your citation gap. They're your highest-leverage off-site targets.

If your own domain rarely appears, list the prompts and ask of each one: which page of ours directly answers this? Often the honest answer is "none".

From MentionX data: citations are a long tail

Method. We ran read-only aggregate queries over completed MentionX runs from 22 August to 29 October 2026 (UTC): 2,154 runs across 18 projects, mostly B2B software, six engines via API with web search or grounding on. We counted returned citation URLs by domain. Brands are anonymised.

How to read it. No single domain dominated. The most-cited domains each had about 1%, and most citations went to vendor sites, niche publications and documentation specific to each category. Advice like "get on Reddit and G2 and you'll be cited" doesn't hold up across these categories. What matters is which domains get cited for your prompts, and only your own baseline can tell you that.

We also compared unbranded answers with and without citations. Answers with no citations mentioned the tracked brand 25.4% of the time (n=682). Answers with citations mentioned it 26.9% of the time (n=1,002). An answer without links can still recommend you, so don't analyse only the answers that have sources.

Caveats. The data covers 18 projects. Citation counts depend on the engine and API settings. Consumer apps may show different sources.

Putting the cuts together

Each cut on its own can mislead. Together they usually point to one cause. Ask three questions in order:

1. Where is the gap? The engine cut and the intent cut give you a cell, for example "ChatGPT, problem prompts".
2. Who wins that cell? Look at the competitor columns for that cell only. Often one competitor takes most of it.
3. What evidence carries the winner? Use the source cut for that cell. Which pages are cited in the answers where the winner is recommended?

The answer to the third question is usually your action. If the winner's own guide is cited, you need a better page. If a review site or a roundup is cited, you need to be on it. If nothing is cited, the engine is answering from what it learned in training, and you're looking at slower work: consistent third-party coverage over months.

Our view: teams that skip question 2 end up chasing the whole market. Teams that answer all three usually find that one or two pages are doing most of the damage.

Accuracy errors

Log every wrong fact: outdated pricing, retired features, wrong markets, wrong integrations. For each:

1. Quote the wrong claim and note which answers repeat it.
2. Find the likely source. Check the cited pages first, then your own older pages, then major listings.
3. Fix your own page first, and date it ("Pricing as of November 2026").
4. Ask the third party to update, and log the request.

Accuracy findings jump the queue. Our view: a wrong price shown to a buyer who already knows your name costs more than a missed mention for one who doesn't.

Signal or noise?

Use the rules from Lesson 6:

• The gap is bigger than the margin of error for that cell.
• It holds across runs, and ideally across more than one engine or several prompts.
• It still holds in the next scan.

A single screenshot is an example, not a trend. Use screenshots to illustrate findings, never to prove them.

Writing findings

Use four lines:

Finding: What's happening, in one sentence with a number and n. Evidence: Rows, prompts and two or three example answers. Likely cause: Your best explanation, marked as a hypothesis. Action: One specific next step, with an owner.

Weak: "Our ChatGPT visibility is low." Strong: "We're named in 2 of 54 ChatGPT answers to problem prompts (Tool A: 31 of 54). Evidence: rows P11–P16, three example answers linked. Likely cause: our duplicate-payment guide is cited, but never names our product. Action: add a product section to the guide; content lead, by 15 November."

Worked example: Ledgerline's three findings

Illustrative example. Ledgerline, its competitors and every number below are invented.

From Baseline v1 (180 unbranded answers, 54 branded), Priya wrote:

1. Problem prompts.

• Finding: Ledgerline is named in 4 of 54 problem-prompt answers. Tool A is named in 29.
• Evidence: Prompts P11–P16, every engine.
• Likely cause: Ledgerline's duplicate-payment blog post is cited 7 times, but never names Ledgerline's own detection feature.
• Action: Rewrite the post with a product section.

2. Comparison sources.

• Finding: On comparison and alternative prompts, the three most-cited third-party pages don't list Ledgerline.
• Evidence: 22 of 54 answers cite at least one of them.
• Likely cause: No profile or a thin one on two review sites, and Ledgerline is missing from one roundup article.
• Action: Complete the profiles and ask the roundup author for inclusion.

3. Pricing accuracy.

• Finding: 11 of 54 branded answers quote the 2024 entry price.
• Evidence: 8 of those 11 cite the same review page.
• Likely cause: That review page.
• Action: Request an update, and add a dated pricing summary to Ledgerline's site.

She didn't write a finding about the 11-point gap with Tool B. It's flagged to recheck next scan.

Mistakes when reading results

• Reporting the overall rate only. It hides the intent and engine where the problem sits.
• Treating citations as the whole story. Answers without sources still name brands.
• Explaining every dip. Small cells move by chance. Only explain changes that pass the signal tests.
• Writing a cause as a fact. "Likely cause" is a hypothesis until a fix changes the numbers.
• Ignoring refusals. A cluster of "I can't recommend vendors" answers on one engine changes what you can expect from it.

Prioritising and presenting

Rank findings by business impact (accuracy first, then comparison and best-of, then problem and learn) and by effort. Take three into Lesson 8.

For stakeholders: open with one sentence on where you stand, show the intent table, then the three findings. Link the raw answers. For a recurring format, use the monthly report template.

Exercise

1. Run the data checks, then make all three cuts on your baseline.
2. Read your five worst answers in full.
3. Write three findings in the four-line format, including at least one about a source.
4. Ask a colleague which one they'd act on first, and why.

Quiz

1. You're present in Perplexity but absent in ChatGPT search. First check? Whether OAI-SearchBot is blocked by robots.txt or bot protection.

2. Cited on problem prompts but not named. What does it suggest? Your content is used as evidence, but it isn't connected to your product.

3. In MentionX data, roughly what share of all citations did the most-cited domain get? About 1%.

4. Is a single screenshot evidence of a trend? No. It's an example.

5. What are the four lines of a finding? Finding, evidence, likely cause, action.

Next: Lesson 8: Your first 30-day action plan.

Sources

• OpenAI, Overview of OpenAI crawlers
• Google Search Central, AI features and your website
• Google Search Console Help, Performance report
• MentionX aggregate scan data, 22 August to 29 October 2026 (UTC), 2,154 engine runs, 18 projects. Anonymised and read-only.