How Method Pages and Experiment Pages Split the Work: One Hard Rule for Content Architecture

What problem does a method page solve for the owner, and what problem does an experiment page solve?

A method page tells people what to do; an experiment page tells people why a conclusion holds. Reading “How to monitor AI citations,” an export business owner wants to know what materials to hand over and what results to look at; reading an observation record, that same owner asks who the model was, how the question was asked, where the original text is, and whether there were failed samples. When the reader’s next step differs, the pages should divide the work.

I treat this rule as an editorial judgment, not a ranking formula. The project’s competitor breakdown lists it as A08: definitions explain the method; experiments disclose data, reproduction, and limitations. This is an actionable recommendation distilled from how pages are organized, and the research itself has not proven that splitting into two pages increases rankings or inquiries.

If you are organizing a research section, you can start with the method for organizing an SEO and GEO topic map to list reader questions, then sort out which ones need long-term explanation and which must rest on one specific observation. Do not first decide how many articles the site should have and then force roles onto ready-made titles.

Where should a method page stop?

The skeleton of a method page should be reusable across experiments. For AI citation monitoring, it has to explain how to fix the questions, how to reduce context contamination, how to preserve the full answer, and how to distinguish a brand being mentioned, a page being cited, and inclusion in a vendor list. If readers can follow it to prepare one record, the page has done its main job.

Judgment criteria are more useful than piling up terminology. For example, “your own domain appears” cannot be written directly as “the brand was recommended”; a company name showing up in a list does not mean the model verified service capability either. Put these boundaries on the method page, and later experiments can reuse them instead of re-explaining them every time.

A method page may include one real example to aid understanding, but the example only serves to illustrate. Do not move the entire round’s result table, every model’s answer, and all revision logs into it, or readers will not be able to tell which parts are long-term rules and which hold only for that round of observation. For a fuller conceptual breakdown, link to the difference between monitoring mentions, citations, and traffic.

Can you write a method page before you have any experiments?

Yes, but you have to make clear that it is an operating plan awaiting verification. You can propose practices such as fixing questions, recording sources, and repeating observations, but you cannot write “we have proven it works.” I separate “we recommend recording it this way” from “we have already done this,” and add evidence links once real results come in.

Above all, do not pad it with unverified client cases just to make the method look mature. If all you have right now is public material and your own reasoning, state the basis clearly. A lack of testing is a state of information, not a blank that must be filled with a story.

Which missing pieces disqualify an experiment page as evidence?

First of all, an experiment page has to let others know what you actually tested. I record the exact question wording, language, date, model, and runtime environment, and state which conditions were fixed and which were not controlled. If you only show a chart of “mentions rising,” readers cannot tell whether the questions changed, the labeling changed, or the answers were fluctuating all along.

The raw record should trace back to a specific answer. The summary table helps reading; the original text supports verification; the two complement each other. If a cited URL merely appears in a source list, keep that identity and do not relabel it as “actively recommended by the AI” in the table. For a record format, see the fields an AI visibility record should keep.

Limitations should sit alongside the results. If a contamination check was not run, write that it was not run; if there is no inquiry data, write that citations have not yet been shown to generate inquiries. Hiding limitations on a distant explanatory page strips a shared result table of its necessary context. I would rather the results look conservative and still let readers restate them correctly.

Keep failed samples too. They help us judge whether the method is too broad and whether a flattering change has another explanation. If you keep only the answers that hit, readers cannot see the prompts where the brand did not appear and therefore cannot understand the scope of observation. On the relationship between original text and sources, read more at how to trace AI citation sources one by one.

How does the fixed GEO panel demonstrate this division-of-labor rule?

Our fixed panel’s work order provides an example well suited to illustrating the division: 24 fixed prompts, each run in 3 independent sessions, producing 72 answers. The baseline was September 17, 2026, and the follow-up observation was October 12. The tool used was dsh (DeepSeek) headless, which does not represent other AI platforms. Here “independent sessions” describes the run method; mirrored-language prompts and repeated prompts cannot be treated as mutually independent statistical samples on that basis.

Under the same criteria, strict mentions went from 3/72 to 6/72, and vendor-identity name mentions went from 0 to 2. The method page explains how “strict mention” and “vendor identity” are judged; the experiment page shows the specific prompts, original text, dates, and labels. That way the numbers have a clear home, and the definitions do not drift along with the results.

This round of records also corrected a false positive: a third-party domain’s path contained a similar string and was mistakenly identified as our own brand. Leave the reason for removing the false positive on the experiment page so readers know what correction the criteria went through. The method page can add a “verify URL ownership” rule on that basis, but it does not need to copy the whole original answer.

The environment description is evidence too. That round only re-ran the third segment of the canary check, so it cannot be written as the whole clean-versus-contaminated environment comparison having been redone; “zero contamination” also lacks supporting detection code. These limits come from the audit addendum to the work order and must not be deleted when polishing the article. The change in results includes sampling noise, and it alone cannot be attributed to a particular content edit.

If you write the experiment as “after we adjusted the content, the AI started recommending us,” you have taken a step the evidence does not cover. I explain further how to preserve this boundary in a research report in the method for reviewing a GEO project without inventing cases.

How should the two page types interlink without duplicating each other?

Interlinks should follow the reader’s question. When the method page discusses labeling standards, link to a real labeling record; when the experiment page discusses environment controls, link to the corresponding method. Anchor text should make clear what the other page can answer — do not just write “click here,” and do not turn an ordinary citation into a “success story.”

I first check whether the target page already carries the same question. If an article explaining the monitoring method already exists, I prioritize updating its evidence entry point; a new experiment can record its own conditions and results independently. Only when the judgment the reader has to make is clearly different do I consider creating a new method page. Swapping the title for a synonym is usually not enough reason to add a page.

One complete result table should have a clear primary home. Other pages cite the necessary conclusions with the date and a link. That way, after the results are revised, you do not have to hunt for old numbers across long articles copied from one another. For the specifics of internal linking, link to organizing topic clusters and descriptive anchor text.

What do you do when an experiment overturns the original method?

Keep the old experiment’s actual conditions while revising the method page. State which judgment is no longer reliable, what the counterexample is, and how to check in the future. The old record still answers “what happened then”; the new method answers “what we recommend doing now.” Wiping the old result into the new criteria destroys the timeline readers use for verification.

When results are updated, what should change and what should stay?

New observations go into the experiment record first; the method is updated only once a change to the judgment criteria is confirmed. If the model’s answers changed while the prompts and extraction rules did not, record the change for that round; if a domain-matching false positive is found, revise the labeling rules and state which historical records need re-checking.

At acceptance, I ask two separate questions: can the method page let someone prepare a record that can be checked? Can the experiment page let someone verify the results and their limits? If the former lacks operating conditions, add conditions; if the latter lacks raw material, add material. You cannot substitute extra length for missing evidence.

For an export business owner, add one more commercial question: what purchasing or marketing decision can this content help you judge? The method lets you evaluate a monitoring plan; the experiment lets you evaluate how much a particular report can be trusted. If there is no data yet on inquiries and closed deals, leave them unproven. Readers can find related observations at the research records entry point and learn who compiles the records from the author and research positioning note.

What common questions come up when applying this rule?

Do method pages and experiment pages have to use different directories?

Not necessarily. What matters is that the title, opening, and interlinks make the page’s role clear and give raw records and explanations their own primary homes. Directories are only an auxiliary way of organizing.

Can an experiment page explain basic concepts?

It can briefly explain concepts that affect the reader’s understanding, then link to the method page. Do not let each round of experiments re-expand into a full tutorial, diluting the conditions, results, and limitations.

If one result improves, can you write that the method works?

You cannot write that directly. You have to check the criteria, the environment, other changes, and sampling noise. Round-to-round change in the fixed GEO panel is only an observation and is not enough for a causal conclusion.

If an old result turns out to contain a false positive, do you delete it or keep it?

Keep a traceable revision note, correct the current conclusion, and state the cause of the false positive and its scope of impact. The old error should no longer serve as evidence, and the correction process should not be hidden either.