AI SEARCH & BRAND DISCOVERABILITY
SITETRAIL CERTIFIED
Your learning path: 12 lessons.
Module 03 / Lesson 09 · Read, apply, check
Build an observation panel that another reviewer can repeat and interpret without mistaking a sample for the whole market.
Define the buyer questions, products, modes, language, relevant location and observation schedule before looking for favorable results. Keep branded questions separate from unbranded category questions: asking about Fieldmark by name tests a different condition from asking for maintenance software options. Record conversation context and use a consistent fresh-session procedure where appropriate. Specify the number of planned observations and how you will handle errors. If a product does not disclose a model version or setting, mark it unknown. The protocol should be detailed enough that another team member can repeat the work without selecting whichever conditions make the client look strongest.
Save the exact prompt, complete visible answer, displayed sources, date and relevant settings. Code mentions, their context, identity accuracy and source verification separately. A failed observation is missing data, not an answer in which the brand was absent. For example, twenty-four planned runs with two technical failures produce twenty-two valid observations and two missing ones. Six mentions among those valid observations can be reported as six of twenty-two, with the missing runs stated. That fraction describes the sampled answers under the protocol. It does not estimate the client’s share of all AI conversations, actual audience exposure or market demand.
Repeat the same panel on the agreed schedule and preserve its version. If a new buyer question becomes important, add it to an exploration set or explicitly revise the panel rather than silently replacing an unfavorable question. A changed interface or unavailable mode may break comparability; record the change and consider a new baseline. Have a second reviewer independently code a small sample, discuss disagreements and improve the definitions. Keep the original captures when correcting a classification. Consistency is valuable because a change in your measurement method can otherwise look like a change in the brand’s actual appearance.
Bing’s published AI Performance description includes citation counts and cited-page information for supported experiences; it explicitly distinguishes those measures from ranking or authority. Review current availability and definitions in the platform before reporting its data. A platform report and your manual panel may cover different products, dates and populations, so do not add them into one unexplained total. Use each for the questions it can answer. The panel can reveal wording and identity errors on selected buyer questions, while an available platform report can show a broader defined activity trend. Neither automatically explains why a particular answer changed.
Fictional Fieldmark observation panel from the lesson; not total market share.
| Record | Count | Interpretation |
|---|---|---|
| Buyer prompts | 12 | 8 unbranded + 4 branded; each run twice |
| Mentioned in valid captures | 6 / 22 | 27.3% of this mixed panel, not a market-wide visibility rate |
| Known obsolete-feature description | 1 | A mention can still contain an accuracy problem |
| Next period mode change | New baseline | Do not draw a continuous trend across incompatible conditions |
The supplied case does not split the six mentions by branded/unbranded prompt. A client report should show that breakdown rather than treating the pooled result as unprompted discoverability.
Fix questions, conditions and planned runs.
Keep complete answers and failed-run records.
Separate mentions, citations, accuracy and missing data.
Review comparable panels and record method changes.
Fieldmark’s agency creates twelve buyer prompts and runs each twice in a specified product and mode during a review window. Eight prompts are unbranded and four ask about Fieldmark directly. Two runs fail to load, leaving twenty-two valid captures. Six valid answers mention the company, including one that describes an obsolete feature. The report presents the counts by question group and preserves the error as a separate finding. A second reviewer disagrees with one recommendation classification; both inspect the full answer and revise the coding note. Next month, the product removes the tested mode. The agency records the disruption and establishes a new baseline instead of drawing a confident trend line across incompatible conditions. Its spreadsheet now supports a repeatable service, while making the limits of the observations visible to the client.
Another agency runs prompts until it collects ten favorable screenshots, discards failed sessions and changes the wording between months. It reports a rising percentage without retaining a denominator. When the client repeats the exercise, the results differ and nobody can reconstruct the original conditions. The problem is not simply that AI answers vary. The agency’s own method also varied, so its headline cannot distinguish a real observation from selection choices made during collection.
Repeatability does not mean every answer must be identical. It means the observation procedure, records and interpretation are clear enough to inspect and repeat. Variation is something to document, not conceal. A modest panel with honest limits can support better decisions than a large unexplained score whose prompts, failures and classification rules are unavailable.
Use permitted, proportionate collection methods and avoid entering confidential client information into public prompts. Platform reports and interfaces evolve. Describe the product and reporting population actually observed, preserve missing data and do not present a non-random panel as a representative survey of all users or every AI system.
Do this today · 30 minutes
Design and interpret Fieldmark’s twenty-four-run observation panel.
Use a document or spreadsheet and the fictional case in this lesson. Work with public or synthetic information; no purchase or confidential client records are required.
Before: Fix questions, conditions and planned runs.
After: Review comparable panels and record method changes.
Notes stay in this browser. Download a copy to keep them.
Ask a colleague to challenge the interpretation. Record the evidence, uncertainty, next action, owner and review date before applying the plan to a client campaign.
Before you move on
Your panel planned twenty-four runs, two failed technically and six of the twenty-two valid answers mentioned the client. Next month the product removes the tested mode. Which reporting and comparison approach preserves the strongest evidence for the client?
Next lesson →Bing: AI Performance reporting and its limits
AMEC: Integrated Evaluation Framework
Lesson edition 1.0 · References reviewed 16 October 2026 · Practical exercise reviewed 16 October 2026. Examples are illustrative teaching scenarios.
Sitetrail service · Guest Post Marketplace
Choose from thousands of publications in the Sitetrail Guest Post Marketplace. Compare audience fit, relevance and budget, then select the opportunities that support your client’s campaign.
Build opportunities for relevant referral traffic, credible media presence and useful discovery. Your agency chooses the mix. Use the marketplace to support a deliberate outreach strategy, and explain the nature and terms of each selected opportunity clearly to your client.
Publication availability and terms vary. Evaluate each opportunity on its own merits; no ranking, coverage or AI citation outcome is guaranteed.
Clear evidence. Useful context. Better decisions.
Use the worksheets to turn each lesson into a real decision. Build a credible campaign as you learn from buyer questions, original evidence and repeatable observations.
Course outline · Free assessment · Privacy
No account is required. Progress, first-attempt lesson answers, worksheets and assessment results are stored in this browser when storage is available. They do not sync across devices. Download copies to retain them; clearing browser data removes local records.
Your free trial starts automatically. No license key is needed yet.
