From Measure to Re-Measure: A 4-Week GEO Closed-Loop Case Study

The company and figures in this article are a composite scenario built from several engagements, not a named customer. Every number is an illustrative value meant to show how one turn of the loop works, not a real measurement. What you should watch is not the numbers themselves but the order in which the loop turns.

Raising your citation rate isn't hard in principle. Answer the question directly, attach evidence, standardize your entity naming, then measure and measure again. Written out as a list, it looks crystal clear. Yet when you actually sit down on Monday morning, it's hard to know where to begin. A list alone doesn't tell you which questions to measure, how many pieces to write once you spot a gap, or when to run your next measurement.

So instead of a list of tactics, this article tells the story in chronological order. We follow a fictional B2B SaaS company through a single turn of the closed loop over four weeks, week by week. Let's say this company sells expense-reconciliation software for small businesses, and we'll call it "Reformeal." It's a team that knows SEO well but is doing GEO for the first time.

Week 0: Number one in search, nowhere in the answer

It started with something small. Reformeal's growth lead asked ChatGPT over lunch, "Recommend a tool for automating expense processing at small businesses." The answer laid out three competitors in tidy paragraphs, and Reformeal was nowhere to be found. This despite the fact that putting the same keyword into Google puts Reformeal's blog at the top of the first page.

Two reactions are common in moments like this. One is to jump straight into churning out content: "AI must not know us yet, let's write more." The other is to shrug it off: "It probably just got skipped that one time." Both make the same mistake: they skip measurement. They move without knowing whether being skipped on one question is a pattern or a fluke, whether it happens only in chatbots or in search-style answers too, or why competitors get cited.

So the decision Reformeal made in Week 0 was simple. For the next week, we write no content and only measure. Because if you don't know your starting line, you can't even say what improved four weeks later.

Week 1: Lock the baseline into numbers

The first step in measuring is deciding what to measure. Reformeal gathered 20 questions its customers were actually likely to ask. They pulled them from sales call notes, support inquiries, and real search terms in the search console. The key was writing them not as abstract keywords like "expense management," but as complete sentences a person would actually type into a chatbot.

• "How do I get started automating expense reconciliation for a small business?"
• "How to automatically categorize corporate card transactions"
• "Is there an alternative to handling expenses in Excel?"
• "How to organize expenses so they're easy to hand off to an accountant"

Next, they put these questions directly to several generative surfaces. But two kinds of surface have to be looked at separately. Conversational chatbots (ChatGPT, Claude, Gemini, and so on) work differently from search-style answers like Google AI Overview. Chatbots answer by mixing learned knowledge with real-time search, whereas search-style answers synthesize by pulling sources from the top of the search results. That's why the brands cited can differ across the two surfaces even for the same question. Vaguely noting "we don't show up in AI" won't cut it. You have to record which question on which surface you're missing from, because that's what determines your next move.

After a week, a baseline map emerged. For example, it looked like this.

Here, citation rate was defined simply as "the share of measured questions where we appear inside the answer." Sophisticated weighting is a problem for later. At the start, what matters more is whether you can measure the same questions with the same definition, over and over. With the baseline locked at about 15% for chatbots and about 10% for search-style answers, there was now something to compare against four weeks later.

Week 2: Read the gaps and set priorities

Lay out the measurement results and you'll see the gaps come in different kinds. Reformeal sorted its missing questions into three buckets.

1. Questions where only competitors appear. Several competitors show up clearly but we're the only ones missing. This signals that the market wants an answer and our seat is empty, so it's the highest priority.
2. Questions nobody answers well. Questions AI answers vaguely. Competition is weak, so there's plenty of room to claim the spot with one precise answer, making this the most efficient bucket.
3. Questions where we already appear. Defend them, but this isn't where to focus right now.

Reformeal picked five questions from buckets 1 and 2. There's a reason for narrowing it to five. The point of the closed loop isn't to crank out lots of content, but to confirm a causal link, whether content leads to citations, within a single turn. If you publish twenty at once, then even if citations rise four weeks later, you can't isolate what worked. To read the cause of a change, you have to change few variables at a time.

There was one more thing they examined while reading the gaps: which sources AI pulls from in the answers where competitors appear. Whether it's the competitor's own blog, an industry outlet, a comparison site, or a community reveals the channel a citation travels through. Once you see the channel, you also see where and in what form to place your content so it can ride the same channel. The list of sources attached to an answer becomes the map for your next move.

Week 3: Publish content that fills the gaps

Now you write. But you don't write the way you'd write an ordinary blog post. You structure each of the five questions from measurement so that a single piece answers it from start to finish. Not one broad, shallow piece like "Everything about expense management," but pieces that answer one thing deeply, like "How to automatically categorize corporate card transactions."

Each piece applied the principles from our overview article on citation rate verbatim. Rather than re-explain the principles themselves, we'll just point out how they actually played out in this case.

• Question as the subheading, direct answer in the first two sentences. They put the sentence a person would type into a chatbot as an h2, and stated the conclusion first, right beneath it. Background explanation came after. The two sentences have to be self-contained enough for AI to lift them straight into an answer.
• Verifiable evidence next to every claim. They didn't invent statistics that don't exist. Where they had their own data, they disclosed it with the time period and sample; where they didn't, they cited official documents or flagged the figure as a hypothetical example. Invented numbers may win short-term citations, but they get filtered out in the end by cross-checks across models.
• Standardized entity naming. They fixed product names, category terms, and the way English equivalents are written into a single form and applied it identically across all five pieces. If you call the same thing differently from piece to piece, the citation signal scatters across naming variants.
• Structured data that matches the body. They added FAQPage schema to FAQ-style pieces, but never filled the markup with content that isn't in the body. Write the body well in a question-and-direct-answer structure, and the schema follows almost automatically.

They didn't pile all their publishing in one place, either. Matching the citation channels read in Week 2, they kept the core pieces on their own site while also accurately reflecting some factual information on external surfaces that AI frequently references. For example, checking factual entries in an industry wiki and correcting errors in directory listings, not slipping in ads, but putting accurate information where it belongs.

There's a common impatience here: the urge to re-measure the very next day after publishing. It takes time for generative engines to crawl new content and factor it into their candidate pool, so measuring right after publishing tends to surface only noise. That's why Reformeal didn't re-measure right away. They waited a week.

Week 4: Ask the same questions again

There's one ironclad rule for re-measuring. Ask the same questions, on the same surfaces, with the same citation definition as in Week 1. Change the questions or the definition, and you can no longer tell whether a change came from the content or from how you measured. Only when you can compare against the baseline does the closed loop finally mean something.

For example, the results came out like this.

Appearances emerged in the areas worked on among the five questions. The chatbot surface moved more, while search-style answers were slower to change. Search-style answers appear to have a lag in reflecting changes, since they pass through an extra step of search indexing and source selection. Had the two surfaces not been looked at separately, that difference would have been buried as a flat "rose about 10pp."

What matters isn't the size of the numbers but the cause and effect you read out of them. Here's what Reformeal learned on this turn.

• The two pieces rewritten in a question-and-direct-answer structure showed up clearly in chatbots, so we scale this pattern up.
• The questions where external-surface information was corrected were reflected late in search-style answers, so we account for the lag and check again at the next measurement.
• One piece showed no change despite being published; the direct answer may not have been self-contained enough, so we'll refine the structure on the next turn.

One turn isn't the end

The real asset Reformeal gained over those four weeks isn't a citation-rate number. It's a way to decide the next content from data rather than guesswork. With the patterns that worked (question-and-direct-answer structure), the patterns with a lag (external-surface corrections), and the piece that didn't work (where to refine the structure) all separated out and visible, the next turn can start on top of this learning.

So the closed loop isn't a campaign you run once and finish. It's closer to an operating rhythm. Models keep changing and competitors are doing the same thing, so a question you appeared on this month might drop you next month. The shorter the cycle (weekly, for example), the faster you learn and the faster you catch changes.

Compress the 4-week closed loop into a table and it looks like this.

Running this loop by hand is possible, but it quickly becomes overwhelming. There are more than twenty questions, chatbots and search-style answers have to be looked at separately, and the re-measurement cycle comes around every week. To record appearances by surface with a consistent definition, connect gaps to content, and re-measure under the same conditions all in one place, you eventually need the help of a tool. Still, the principle to remember comes before any tool: citation rate rises not from writing more, but from measuring first and validating one variable at a time. NUDGEO helps you check that starting line and measure again under the same conditions.