Share of Model Checklist: The Essential Best Practices
Share of model — how often your brand shows up, gets cited, or gets recommended inside AI answers — is easy to talk about and surprisingly easy to measure badly. Most of the programs we’ve seen fail quietly: a sloppy prompt set, no scoring rules, and a spreadsheet nobody trusts after month two. This is the audit list. Run your program against it before you present a single chart to a client or a boss.
We’re not re-explaining the full measurement workflow here — that’s covered elsewhere in this series. This piece is the checklist you pin to the wall: what a defensible share-of-model setup actually requires, area by area, and what to check when your numbers start improving your score for real reasons instead of measurement noise.
Prompt-Set Design
- Prompts are written in the customer’s language, not internal jargon or brand-speak
- Set includes navigational, comparison, “best of,” and problem-first prompt types
- Prompts map to actual buying-stage intent (awareness, comparison, decision)
- Set size is large enough to smooth out single-answer variance (dozens, not five)
- Prompt wording is version-locked so re-runs are truly apples-to-apples
- A subset of prompts intentionally excludes your brand name, to test unprompted recall
The prompt set is the instrument. If it’s biased toward queries you already win, your share-of-model number is just a mirror telling you you’re pretty. The point of an AI-search-era measurement program is to find where you’re invisible, not to confirm where you’re already strong.
Version-locking matters more than people expect. Change a prompt’s phrasing mid-quarter and you’ve broken the trend line — you’re no longer measuring visibility change, you’re measuring wording change. Treat the prompt set like a survey instrument: once it’s fielded, it doesn’t move until the next planned revision.
Platform Coverage
- At minimum, coverage spans ChatGPT, Gemini/AI Overviews, Perplexity, and Claude
- Each platform is tested on its native surface, not a third-party aggregator only
- Logged-in vs. logged-out (or memory-off) behavior is noted, since personalization skews results
- Voice and app-based variants are flagged separately if they’re materially different from web
- Platform mix is reviewed periodically, since usage share among tools shifts
Every one of these engines pulls from different sources, weights citations differently, and updates on its own schedule. A brand can be well-represented in one and functionally absent in another. Reporting a single blended “AI visibility score” without platform breakdown hides exactly the information a marketing team needs to act on.
Logged-in state deserves its own line item because it’s the most common source of false confidence. If your own accounts have search history full of your brand, of course the assistant mentions you — that’s personalization, not organic share of model. Test from a clean or logged-out context whenever the platform allows it.
Tagging and Scoring Rules
- Written definitions exist for “mention,” “citation,” and “recommendation” before scoring starts
- Sentiment is scored on a fixed scale (e.g., positive/neutral/negative), not gut feel per reviewer
- Competitor mentions in the same answer are logged alongside your own
- Position within the answer (first mentioned vs. buried) is captured, not just presence/absence
- A second reviewer spot-checks a sample of scored answers for consistency
This is the area most programs skip, and it’s the one that determines whether your numbers mean anything. “Mentioned” is not the same as “recommended,” and treating them as interchangeable inflates the score without telling you anything useful about buying influence. Write the rules down before the first scoring session, not after you notice two people scored the same answer differently.
Position matters because AI answers aren’t read the way a ranked list is read, but they’re not read uniformly either. A brand named first, with a real explanation attached, is doing different work than a brand tacked on in a closing list of five options. Capture that distinction or you’ll flatten a meaningful signal into a single checkbox.
Cadence and Tracking
- Testing cadence is fixed (e.g., monthly) and documented, not run whenever someone remembers
- Each run’s date, platform versions, and any known model updates are logged with the data
- Historical runs are archived in full, not overwritten, so trend lines survive
- A baseline run exists before any optimization work started, for honest before/after comparison
- Sudden swings trigger a re-run before being reported as real movement
AI answers are non-deterministic and the underlying models change without notice. A single run is a snapshot, not a trend. The value in share-of-model tracking comes entirely from consistent, repeated measurement over time — anything less is closer to a poll than a metric.
Logging model updates next to your data isn’t busywork. When a score jumps or drops sharply, the first question should be “did the model change” before “did our content change.” Skip that logging step and you’ll eventually credit or blame a content update for something a platform did on its own.
Competitive Benchmarking
- A fixed competitor set is tracked alongside your own brand on every prompt
- The competitor list is reviewed periodically for new entrants, not set once and forgotten
- Share is reported as relative standing (your mentions vs. category mentions), not just raw counts
- Category leaders and category challengers are both included, to see the full spread
- Notes capture *why* a competitor won a given answer — cited source, review volume, etc.
A share-of-model number in isolation tells you almost nothing. Twelve mentions out of fifty prompts sounds fine until you learn a competitor got thirty-five. Context is the whole point of “share” — it’s a competitive metric, not a vanity one, and it should be treated with the same rigor as a market-share study.
The “why” notes are what turn a scoreboard into a strategy. If a competitor consistently gets cited because of a specific comparison page or a heavily reviewed product listing, that’s an actionable finding. Recording only the win/loss count and skipping the reason wastes the most useful part of the exercise.
Improving Citations and Mentions
- Published content directly answers the exact question phrasing used in your prompt set
- Claims and data points are attributable to a clear, named source on the page
- Author identity and real experience are visible, not hidden behind generic bylines
- Structured data (FAQ, article, organization schema) is present and accurate
- Third-party mentions — reviews, comparisons, forum threads — are actively earned, not just owned content
- Outdated or contradictory claims across your own site are found and reconciled
Generative engines lean heavily on content that reads as a clean, citable answer: a direct claim, a clear source, a specific number or named example. Vague marketing copy rarely gets pulled into a synthesized answer because there’s nothing precise to extract. This is where E-E-A-T stops being an abstract framework and becomes a practical lever — named, experienced authors and verifiable claims are exactly what these systems are built to surface.
Owned content is only half the picture. Independent mentions — a comparison written by someone else, a well-answered thread, a review site’s writeup — carry weight precisely because they aren’t self-published. Treat earned mentions as a deliberate target, not a side effect of PR.
At SEO University, run by Salterra Digital Services, we treat this the same way we treat traditional link earning: identify the third-party surfaces your buyers already trust, and invest in being genuinely worth citing there — not just optimized to be found.
Data Hygiene
- Raw answer text is saved with each score, not just the numeric tally
- Source URLs cited within AI answers are recorded and checked for accuracy
- Scoring spreadsheet or database has a single owner responsible for schema consistency
- Duplicate or malformed entries are cleaned before any report is generated
- Access and edit history are logged so a bad edit can be traced and reversed
None of the previous checklists matter if the underlying data is a mess. Saving raw answer text, not just a score, is what lets you re-audit a surprising result six months later instead of trusting a number you can no longer explain. This is unglamorous work, and it’s exactly the work that separates a measurement program people trust from one they quietly stop believing.
Frequently Asked Questions
How many prompts do I need for a reliable share-of-model score?
There's no universal number, but a set too small to smooth out normal answer variance will produce a score that swings for reasons that have nothing to do with your visibility. Err toward more prompts across more intent types rather than fewer, and keep the set stable once it's fielded so trends stay comparable.
Should I count a citation the same as a recommendation?
No. A citation means your content was used as a source; a recommendation means the assistant actively suggested your brand as a choice. They indicate different levels of influence and should be scored on separate lines, even if you later combine them into a summary metric.
How often should share-of-model be measured?
A fixed, documented cadence — commonly monthly — is more valuable than frequent but irregular checks. Consistency is what makes the trend line trustworthy; sporadic testing makes it impossible to tell a real shift from ordinary answer variance.
Does being logged in to an AI platform affect the results?
Yes, often significantly. Personalization and account history can surface your brand more (or less) than a neutral user would see. Whenever the platform allows it, test from a logged-out or clean context and note when that isn't possible.
What's the fastest way to improve a low share-of-model score?
Start with the checklist under Improving Citations and Mentions: make claims specific and sourced, put real named expertise behind the content, and pursue independent third-party mentions rather than relying only on owned pages. Generative engines favor precise, attributable, well-structured answers over generic marketing copy.
Can share-of-model replace traditional SEO tracking?
No — treat it as a complementary layer. Search rankings and AI-answer visibility respond to overlapping but distinct signals, and a brand can perform very differently across the two. A complete measurement program tracks both rather than substituting one for the other.
Terry has 30+ years in software and SEO. He’s the founder of Salterra Digital Services and SEO Spring Training, host of the Roundtable SEO Mastermind, and lead instructor at SEO University — teaching the exact tactics his team uses on client work.
Continue learning
Ready to master this?
This guide is one lesson from the Measuring AI Visibility Share of Model course. Get every lesson, framework and checklist — plus the full 38-course catalog — inside SEO University.
Practitioner-focused training across the full digital marketing stack — from technical SEO to conversion optimization and the AI search era. By Salterra Digital Services, since 2011.