Share of Model: The Visibility Metric That Replaces Keyword Rank in AI Search

The Metric That Replaces Keyword Rank

Share of Model is the percentage of AI-generated answers in a category where an engine names or cites your brand, measured across a fixed query set plus every major engine. It replaces keyword rank because the ranked list of links is no longer where answers happen: on a Google results page carrying an AI summary, users click a traditional link in just 8 percent of visits, down from 15 percent without one, per Pew Research Center. Digital Strategy Force measures it with the DSF Answer Presence Index, which composites four inputs into one score: Mention Rate, Citation Rate, Prominence Weight, plus Engine Coverage.

The twist is that Share of Model is not one number but one number per engine, plus a brand's real reach is capped by its weakest engine, not its best. A category-leading presence in ChatGPT beside a zero in Gemini does not average into visible. This is why a keyword-rank report plus a Share of Model report can disagree completely: one counts a position on a page almost no one acts on, the other counts presence inside the answer the user actually reads.

The scale of the shift is what forces the new metric. Google's AI Overviews now reach 2.5 billion monthly users plus its AI Mode passed a billion in a year, with each AI Mode question running triple the length of a traditional search. When one question fans out into a dozen parallel queries, a single keyword position stops describing your visibility at all. The figures below quantify the collapse of rank before the Answer Presence Index names what to measure instead.

Why Position on a Page Stopped Measuring Visibility

Keyword rank lost its meaning the moment the ranked list stopped being where answers are delivered. When an AI summary appears, Pew Research Center measured the click rate on traditional links falling from 15 percent to 8 percent, only 1 percent of users clicking a link inside the summary, plus the session ending on the spot 26 percent of the time, versus 16 percent without one. A number one position still exists; it just sits below the answer most people never scroll past.

The behavior change is not marginal. Bain finds roughly 60 percent of searches now end without a click, plus 80 percent of consumers rely on zero-click results in at least 40 percent of their searches. The traffic that rank used to predict is draining: the Reuters Institute measured Google organic referrals to 2,500 sites falling 33 percent globally plus 38 percent in the United States in a single year, with publishers expecting another 40 percent over three years.

"Keyword rank measured one slot on one page. Share of Model measures presence across every answer in a category, plus presence is capped by the weakest engine, never the average."

— Digital Strategy Force, Search Intelligence Division

The change is structural, not cyclical, because the answer surface is now the default for over a billion people. The stat cards below set the collapse of rank against the rise of the answer, plus every figure points to the same conclusion: visibility has to be measured where the answer is built, not where the links used to be.

What Share of Model Actually Measures

Share of Model resolves into four measurable inputs, not one headline number. The DSF Answer Presence Index composites Mention Rate, Citation Rate, Prominence Weight, plus Engine Coverage, computed per engine before they are aggregated. The four compose multiplicatively, so a zero on any one input collapses the score: a brand named in every answer but never linked, or named on one engine only, does not have presence, it has a fragment of it.

Input 1, Mention Rate: the share of answers in your query set that name your brand at all, whether or not they link it. This is the broadest signal plus the one most tools stop at, because it is the easiest to count. On its own it overstates visibility, since being named fourth in a list is counted the same as being the answer.

Input 2, Citation Rate: the share of answers that link your page as a source, not just mention the name. A citation routes attribution plus traffic in a way a bare mention does not, so Citation Rate separates the brands an engine merely knows from the ones it is willing to stand behind.

Input 3, Prominence Weight: where the mention sits plus how it is framed, because position plus sentiment decide whether a user ever reads it. Raw counting is biased here: models systematically under-cite numeric content by 22.6 percent plus personal names by 20.1 percent, per 2026 research on citation behavior, so an unweighted tally misreads who is actually surfaced.

Input 4, Engine Coverage: how many of the engines a category's buyers use actually name you. This is the input that turns Share of Model into a per-engine measure, because the calculation each engine runs differs, plus a brand strong on one can be invisible on the next. The diagram below shows how the four inputs compose into a per-engine score.

Prominence Weight is the input most measurement programs get wrong, because they count mentions as if the model treats every fact the same way. It does not. The same 2026 work found models over-cite text already flagged as needing a citation by 27 percent while under-citing the exact evidence buyers care about, the numbers plus the named entities. A tally that does not correct for that bias rewards the wrong pages.

Why Share of Model Is Computed Per Engine

A brand's Share of Model is not one number; it is one number per engine, because the engines do not share a retrieval pipeline, an index, or an audience. Perplexity crawls in near real time, ChatGPT leans on a different index, plus Gemini weights its own knowledge graph, so the page that wins one can be missing from the next. Averaging the four into a single score hides exactly the gap that matters.

The divergence shows up in hard numbers. Cloudflare measured Anthropic's crawler fetching 38,000 pages for every visitor it referred back in 2025, against 1,091 for OpenAI plus just 40.7 for Google, a thousand-fold spread in how each engine reads plus returns the web. Audiences split the same way: Pew found 59 percent of US teens use ChatGPT against 23 percent on Gemini, so the engine that owns your category's buyers is the one your score has to weight.

This is the Per-Engine Floor Principle: your real reach is the floor across engines, not the average. A cross-engine citation audit that reports one blended figure will tell a brand it is visible when it is actually dominant on one engine plus absent on the three its buyers also use. The dumbbell below shows what a single average conceals.

Brand A looks healthy on a blended number plus is one query change away from a collapse, because almost all of its presence sits on a single engine. Brand B scores lower on average plus is the safer position, since its reach holds across engines. The table below puts the per-engine differences in concrete terms, so a measurement program knows why it cannot collapse them.

The Five Inputs of a Defensible Measurement Program

A Share of Model number is only trustworthy when five inputs are defined first: Query Surface, Engine Coverage, Capture Cadence, Quality Weighting, plus Attribution Linkage. The DSF Measurement Surface Scorecard names them, because a figure produced without them measures luck, not presence. Leave any one undefined plus the score becomes a vanity number that moves with sampling noise rather than with the market.

Query Surface: a fixed, funnel-weighted set of 50 to 200 buyer questions, versioned over time. It has to reflect how people actually ask, plus AI Mode questions already run triple the length of a traditional search, so a surface built from short head keywords measures a query universe that no longer exists.

Engine Coverage: every engine your buyers use, scored separately, never blended. This is the input that operationalizes the Per-Engine Floor Principle, plus skipping an engine because it is harder to query is how a program reports a strong score over a weak floor.

Capture Cadence: a fixed re-run schedule, because AI answers are non-deterministic plus a one-time check captures one roll of the dice. Presence is a distribution across runs, prompts, plus time, so the cadence is what separates a trend from an anecdote.

Quality Weighting: position plus sentiment applied to every captured mention, so a lead recommendation is not scored the same as a fourth-place aside. Without it, the bias that makes models over-cite some content plus under-cite numbers passes straight into your number.

Attribution Linkage: a join from the score to a business outcome, because AI citations rarely pass referrer data. Without a path to branded-search lift, assisted conversions, or CRM closeback, Share of Model stays a presence metric that cannot defend a budget. The scorecard below makes each input auditable.

Turning Share of Model Into a Number That Moves

Share of Model rises on a short list of levers: structure the model can extract, corroboration it can trust, plus presence on the engines a category's buyers actually use. None of them is a keyword. Structural feature engineering research lifted citation rate by 17.3 percent plus answer quality by 18.5 percent across six engines from structure alone, which is why the metric responds to how a page is built, not how often a phrase appears in it.

A worked example shows the loop close. A mid-market B2B SaaS firm scored a high Mention Rate on ChatGPT plus a near-zero on Gemini, so its blended number looked acceptable. The Measurement Surface Scorecard exposed the floor: no Engine Coverage on the platform its enterprise buyers used. The team rebuilt three pages for extractable structure plus corroborated their key figures, plus within six weeks the Gemini floor lifted, plus the real Share of Model, the floor, moved for the first time in two quarters.

Earned prominence is now a visible lever too. Google's May 2026 launch lets users pin Preferred Sources, with more than 345,000 already selected plus people twice as likely to click one, plus its Highly Cited badge flags the corroborated, primary reporting. The cards below set the levers against the evidence behind each.

What Share of Model Cannot Tell You

Share of Model measures presence, not profit, plus a high score with no attribution linkage is a vanity number. Presence is not traffic: Cloudflare's 38,000-to-one crawl-to-referral ratio is proof that an engine can read you constantly plus send almost no one, so a rising score does not promise a rising visit count. The metric tells you whether the answer names you, never whether that naming paid.

"A Share of Model score with no attribution linkage is a vanity metric. Presence that cannot be tied to pipeline is a number that feels like progress plus funds nothing."

— Digital Strategy Force, Search Intelligence Division

It also cannot be trusted from a single run. AI answers are non-deterministic, plus research probing how models attribute knowledge found attribution mismatches raise error rates by up to 70 percent, so one capture can show a brand named that the next capture drops. The panel below states the three boundaries to hold in view, so the metric is used for what it measures plus not stretched past it.

Hold those limits plus the metric still does the one thing keyword rank no longer can: it counts visibility where the answer is built. When a number one position converts at 8 percent of the clicks it once did, the brands that win are the ones the answer names, plus Share of Model is how that presence becomes a number a team can move. Keyword rank measured the page. Share of Model measures the answer.

FAQ — Share of Model Measurement

Next Steps — Share of Model Measurement

Digital Strategy Force builds Share of Model programs the same way, every time: define the surface, score the inputs, find the floor, then wire the number to revenue.

Digital Strategy Force Answer Engine Optimization runs the DSF Answer Presence Index across your category, finds the floor engine capping your reach, plus ties your Share of Model to pipeline, so you measure the metric that replaced keyword rank instead of the one that no longer counts.