Why I Measure Share of Citation

I measure share of citation because a buyer can be given an answer that knows our name and still be given nobody's link to us. Citation share is our qualifying citations as a percentage of all qualifying citations in the same sampled answers. It is a share of the cited-source pool. A mention in the prose, a ranking, and an impression do not tell me whether the engine selected our page as evidence.

The definition, the formula, and the tracking method are in AI Citation Share Definition Formula and Tracking Method. That page is the deeper method. This essay is how I would use it as a founder: why this is the number I manage, how I read the studies that justify it, and what would make me manage a different number instead.

The arithmetic I refuse to let a dashboard rename

Citation share equals our qualifying citations divided by all qualifying citations in the same sampled answers. Run 100 buyer-intent prompts across ChatGPT, Perplexity, Gemini, and Claude. If those answers issue 320 citations and 45 point at us, citation share is 14.1%. Divide by the answers instead and the result is citation rate. Nineteen of those 100 answers citing us at least once is a 19% citation rate. Both numbers come from one sample. Neither corrects the other. A percentage with no denominator cannot be compared with anything, which is why I want the name of the denominator in the same sentence as the percentage.

A mention is the brand name in the answer with no source link. A citation is the engine attaching a URL as evidence. Citation share counts citations. Share of voice counts appearances. A brand can score well on share of voice and hold zero citation share. The engine knows the name. It does not trust the page enough to use it. That is an authority problem. Calling it a visibility problem sends the team to the wrong work.

I started tracking the gap in 2024, when strong organic rankings and an empty answer were showing up on the same brands. I use the name share of citation for the selection, inside the discipline I coined the same year. The canonical definition is the glossary entry. The operating order around it is Machine Relations.

What the large studies changed about where I spend

Semrush's June 2026 AI Visibility Index analyzed 126 million U.S. AI search prompts across ChatGPT, Gemini, Google AI Mode, and Google AI Overviews (BusinessWire). Being mentioned and being cited are different. On Gemini, the overlap between mentioned brands and cited domains drops as low as 30%. A brand in the answer text is still not cited as a source seven times out of ten. A Columbia Journalism Review analysis found that only 49% of answers across eight generative search tools included any citation at all (CJR). When the answer layer cites less than half the time, each citation I do capture carries more weight than it did in a world where every answer linked its sources.

Clearsight ran 600 product queries through Google AI Overviews across 12 ecommerce categories and logged 4,184 citations (Clearsight). In that ecommerce segment the top 3 cited domains capture a median of 47% of weighted citations and the top 10 capture 78%. I treat that as one segment's result, not as a law of every category I operate in. Across 16,475 monitored answer runs in the September 23, 2026 Machine Relations Index release, the ten most-cited domains in a category hold between 7.0% and 24.2% of that category's citations, and reaching half of a category's citations takes between 49 and 304 distinct domains. Concentration is worth watching. A founder outside a top ten is not invisible. The middle of the market is where a lot of reachable sources sit.

Everything-PR's Google AI Overviews Citation Source Index, published June 2026, found YouTube leading AIO citations at 20.9% citation share, with the top 1% of domains, roughly 12 sites, capturing 47% of citations. The median cited page is 14 months old. Pages with schema markup are cited 2.3 times as often. The relationship between organic rank and AIO citation has weakened: only 38% of cited pages also appear in the top 10 organic results for the same query, down from 76% in mid-2025. I stopped using a top-10 ranking as evidence that we will be cited. The 38% is the overlap. It is not a plan.

The predictor I would rather buy is branded mention velocity: how often a domain is mentioned across the open web in a month, weighted by source authority. In the Clearsight study the correlation with weighted citation share was 0.74. Ahrefs Domain Rating correlated at 0.31. Several domains with DR under 50 outperformed domains with DR over 80 in the same category. The model is not picking by raw domain authority. It is picking by how often the entity appears in fresh, relevant pages. Of the 4,184 citations, 217 came from brand-owned domains, roughly 5%. Publishing only on our own blog will not move citation share at scale. I need third-party coverage from sources the engine already trusts. A clipping report with flat mention velocity is a report that the placements are not entering the retrieval path. A DR of 80 with no independent publisher writing about us in the category is a backlink profile the engine is free to ignore.

I have written the earned-media side of that in how PR affects AI search visibility. The version I use in a budget meeting is shorter. Mention building on category-relevant publications is the work. Link building is a different work, and the 0.31 correlation is why I do not lead with it.

Engines do not share a list, so one engine is not the number

Foglift's second-quarter 2026 benchmark tested 75 brand-neutral buyer-intent prompts across 25 verticals on five engines: ChatGPT, Claude, Gemini, Google AI Overview, and Perplexity. Cross-engine Jaccard similarity was 0.18. Any two engines share less than a fifth of their cited domains for the same queries.

Semrush's 126-million-prompt study found the same split at a different scale. ChatGPT cites an average of 15 sources per response. Gemini cites an average of 3. That is a fivefold difference in citation density on the same buyer's question. Only 36 brands held top-100 visibility across all four platforms in every month of the study. A brand with 25% citation share on Perplexity and 3% on ChatGPT has a structural gap. It does not have a measurement error. Each engine's list reflects different retrieval, different source preferences, and different weights on freshness, authority, and structure.

Google AI Mode, launched in 2026, pulls from Google's index and still selects differently from AI Overviews. Ahrefs compared 730,000 AI Mode and AI Overview responses to the same queries and found that only 13.7% of the URLs they cite overlap (Ahrefs). Two Google products, one query, two citation lists. I report per engine. An aggregate with the engines hidden is how a strong Perplexity month disguises an empty ChatGPT month.

The free baseline, and the way it lies

On June 16, 2026, Microsoft launched Citation Share inside the Bing Webmaster Tools AI Performance dashboard (Bing's announcement). It is the first time a major engine has given publishers first-party citation share at no charge. If a query generates 40 citations and the site captures 6, citation share for that query is 15%. Three companions shipped with it. Intents classifies grounding queries as informational, commercial, research, navigational, learn and solve, creation, or local. Topics groups queries the way the system reasons across them. Compare overlays a prior period so a content change can be seen against the previous window.

The scope note matters more than the feature list. Bing's number covers Copilot and Bing. It does not reflect Google AI Overviews, Perplexity, or ChatGPT. For those I still need a fixed prompt set, run on purpose, or a tool that states its own method. Semrush, Otterly, Profound, AthenaHQ, and Ahrefs have all added some form of AI citation tracking. I pick one method and keep it. In early July 2026 Microsoft confirmed that large citation surges site owners saw from June 1 were data backfilling, not a visibility breakthrough (PinMeTo). I use late June as the clean baseline and I treat a spike on June 1 as housekeeping.

Before that dashboard existed, every founder measuring this was building a monitor or paying for one. A zero-cost baseline on one engine family is a real change. It is not cross-engine share.

The bands I use, and the bands I refuse to universalize

There is no universal benchmark, because concentration varies by category. I still keep bands so a meeting has words for a number.

Above 35%, I treat the answer layer in that query set as ours to defend. From 20% to 35%, we are consistently cited and the work is expanding coverage and making the strong clusters harder to take. From 10% to 20%, we are in the game and not controlling it, and the per-engine cut will show where it falls off. Below 10%, the layer is not selecting our pages. Either the engine does not know who the company is, which is entity work, or it does not trust the page enough to cite it.

Semrush's concentration figures tell me how hard the band is to enter. In News and Media the three most visible brands account for 82.9% of category visibility. In Consumer Electronics the top three are 76.9%. In Finance they are 41.4%. A concentrated category is harder to break into. A distributed one, Finance in this cut, is where a focused brand can gain ground. I do not borrow the News band as a target for a finance company.

On a 34-query commercial set in a July 2026 snapshot, 27 of 34 queries produced a win and 184 of 813 citation slots were captured, 22.6% citation share. That number moved because of what was published and where it was placed, not because the volume of pages went up. I use it as an example of a set with a denominator, not as a promise to the next company. Everything-PR's standing Citation Share Index uses roughly 28 entities per category, 62 buyer-intent prompts, and five engines, and it keeps finding the same concentration: a small number of sources take the majority in every category studied.

Clearsight also found that 68% of citations came from content published or substantially updated within the prior 12 months. Freshness decays. A quarterly refresh of the strongest pages is a lever I will fund. A controlled test of seven citation-tracking tools on the same domain over 15 days found an 8.2-fold gap between the lowest and the highest citation count. The tools count different things. Some count mentions, some count linked citations, some use different prompts, some sample at different frequencies. The trend inside one method is the signal. The absolute number across tools is a argument about definitions.

How I run it without turning it into a second SEO report

I define 50 to 100 buyer-intent queries. I skip "what is our company." I want the questions where someone is deciding, comparing, or looking for evidence. I run them on ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews. For Bing I use the native metric and I remember what it does not cover. I count linked citations, not bare mentions. I report per engine and in aggregate. Aggregate is position. Per engine is where the next dollar goes. I track monthly for a baseline and weekly when a campaign is in market.

Metric	What I am actually seeing	Where it stops
SEO ranking	Position on a results page	Silent on answer selection
Share of voice	How often the name appears	Counts appearances with no URL
GEO visibility	Presence inside a generated answer	Mixes a mention with a citation
Citation share	Share of the cited-source pool	Needs many engines, and tools can disagree eightfold

Citation share sits at the narrow end of that stack, which is why I manage it. I have watched the other three look healthy while the cited-source list never included the company. The engine knew the name. It did not use the page. I pair the metric with entity work so the engines know who the page belongs to. Citation share without entity clarity, or entity clarity without a cited page, is half the system.

Which concentration number I would show

I would not put Clearsight's 47% and the Index's 7.0% on the same slide without the denominators. They answer different questions, and both are in the record I am willing to use.

Clearsight's median, top 3 domains at 47% of weighted citations and top 10 at 78%, is one ecommerce segment: 600 product queries, 12 categories, 4,184 citations, Google AI Overviews. Everything-PR's June 2026 index is a cousin of that picture. YouTube at 20.9% of AIO citation share, the top 1% of domains, roughly 12 sites, at 47%. If I am briefing an ecommerce catalog, those are the slides. The median cited page at 14 months, schema cited 2.3 times as often, and only 38% of cited pages also ranking in the organic top 10, down from 76% in mid-2025, belong on the same slide. They tell a merchandiser that the organic report and the citation report have come apart.

The Index figure is the one I would show for a category that is not that ecommerce segment. Across 16,475 monitored answer runs in the September 23, 2026 release, the ten most-cited domains hold between 7.0% and 24.2% of a category's citations, and half the citations take between 49 and 304 domains. A founder outside the top ten is not invisible. The middle is where reachable sources live. Semrush's category spread says the same thing in brand-visibility language: the top three are 82.9% in News and Media, 76.9% in Consumer Electronics, and 41.4% in Finance. I would be embarrassed to set a 35% "dominant" band as the goal in a category whose leaders share the way Finance does, and I would be naive to treat 10% as a foothold in News. The bands are words for a meeting. The category is the benchmark.

Cross-engine rarity is the third slide, and it is why I will not accept a blended share. Foglift's Jaccard of 0.18, on 75 prompts, 25 verticals, and five engines, means two engines share less than a fifth of cited domains. ChatGPT's 15 sources against Gemini's 3 is a fivefold difference in how many slots even exist. Only 36 brands held a top-100 position on all four Semrush platforms in every month of that study. Google AI Mode and AI Overviews sharing 13.7% of cited URLs, in Ahrefs' 730,000-response comparison, means I cannot even treat Google as one engine. A 25% Perplexity share beside a 3% ChatGPT share is the gap I would fund. An average of those two is how I would hide it.

The owned-versus-earned slide is the one that changes the work. Two hundred seventeen of 4,184 Clearsight citations, about 5%, came from brand-owned domains. Mention velocity correlated at 0.74. Domain rating correlated at 0.31, and domains under DR 50 beat domains over DR 80 in the same category. Sixty-eight percent of the citations came from pages published or updated in the prior 12 months. The dollar goes to third-party mentions on category sources, and to refreshing the pages that are already being selected. It does not go to another DR campaign or to a higher volume of posts on our own blog.

Bing is the free way to practice the habit on one engine family, with the limits written on it. Since June 16, 2026 the AI Performance dashboard reports citation share for Copilot and Bing. Forty citations with 6 of them ours is 15% on that query. Intents, topics, and the compare overlay are useful inside that fence. The June 1 backfill Microsoft confirmed in early July is why the baseline starts in late June. I will not walk into a meeting with a June spike and call it momentum. And I will not let a Copilot percentage stand in for Perplexity, ChatGPT, or AI Overviews.

If two tools disagree, I do not split the difference. Seven tools, one domain, 15 days, an 8.2-fold gap. I keep one method, name the denominator, and manage the trend. The July snapshot I use as a worked example, 27 wins out of 34 commercial queries and 184 of 813 slots, 22.6%, is a number with a denominator. It is not a target I hand to the next company. It is proof the metric can be computed on a set small enough to run.

The decision

I would not run an AI-visibility budget on share of voice, on domain rating, or on a single engine's dashboard. I would run it on citation share, per engine, on a fixed set of 50 to 100 buyer queries, with the denominator written down. I would treat Bing's free number as the Copilot baseline from late June 2026 forward, and I would not let a June 1 backfill become a victory. I would put the next dollar into third-party mentions on category-relevant sources, because 0.74 beats 0.31 and because only about 5% of the Clearsight citations came from brand-owned domains.

The scope is selection inside sampled answers. It does not say a mention is worthless. A mention is how the velocity predictor moves, and velocity is what predicts the citation. It does not say 35% is a universal trophy. Clearsight's top-3 median of 47% and the Index range of 7.0% to 24.2% for a category's top ten domains are different markets. I compare inside the category and the query set.

I would change my mind if a later multi-engine study, with the method published, showed domain rating predicting citation share as strongly as branded mention velocity, or showed brand-owned pages taking a large share of citations rather than roughly 5%. I would also retire a tool the week its count diverged from a second tool on the same domain with no explanation, given that seven tools already disagree by 8.2 times. The trend inside a named method can survive that. A vendor's absolute number cannot. Until one of those reversals is in front of me, selection is the number I manage, and everything else is context for it.

FAQ

What is citation share, in one calculation?

Our qualifying citations divided by all qualifying citations in the same sampled answers. One hundred buyer prompts that produce 320 citations, 45 of them ours, is 14.1% citation share. Nineteen of those 100 answers citing us at least once is a 19% citation rate. I want both, and I want each one named. A bare percentage is not a metric.

What do I call a good score?

Inside a competitive B2B query set I use four bands: above 35% to defend, 20% to 35% as strong, 10% to 20% as in the game, below 10% as not selected. Clearsight's ecommerce segment put the top 3 domains at a median of 47% and the top 10 at 78%. The Index's category range for the ten most-cited domains is 7.0% to 24.2%, and half of a category can take 49 to 304 domains. I use the band inside our category. I do not paste the ecommerce median onto a different market. Semrush's own spread, 82.9% for the top three in News and Media and 41.4% in Finance, is why.

How is this different from share of voice?

Share of voice counts appearances: mentions, impressions, placements. Citation share counts the times an engine attaches our URL as a source. Semrush found Gemini's overlap between mentioned brands and cited domains as low as 30%. The CJR reading found citations in only 49% of answers across eight tools. A familiar name with no link is recognition. I do not pay a citation program for recognition.

Can I see any of this without a new contract?

Yes, for one engine family. Bing Webmaster Tools has reported Citation Share since June 16, 2026, for Bing and Copilot. Microsoft's backfill from June 1 means I start the baseline in late June. ChatGPT, Perplexity, and Google AI Overviews are not in that dashboard. Google's own AI Mode and AI Overviews share only 13.7% of cited URLs on the same queries. I run a fixed prompt set for those, or I use one tool and keep its method still. I do not average Bing with Perplexity and call it a share.

What moves the number?

Branded mention velocity, correlation 0.74 in the Clearsight study of 4,184 citations, against 0.31 for Domain Rating. Fresh third-party mentions in the category beat raw authority. 68% of those citations came from pages published or substantially updated in the prior 12 months, so I refresh the pages that already get selected. 217 of 4,184 citations were brand-owned, about 5%, so I do not expect the blog alone to carry the share.

Why do two tools give me two shares?

A 15-day test of seven tools on one domain produced an 8.2-fold gap between the low count and the high count. They are not failing at the same measurement. They are measuring different objects. I pick a method, name the denominator, and manage the trend. ChatGPT's 15 sources against Gemini's 3, and a cross-engine Jaccard of 0.18, mean even a perfect tool has to report engines separately. One blended number is how I would hide the gap I most need to see.