Whether I Would Spend to Raise Entity Mass

I would spend the next dollar on cross-domain evidence that makes the company a single, recommendable entity, and I would pause net-new pages until that evidence is above the line these studies describe. Most brands in the datasets are absent from AI recommendations. The absence tracks how many independent domains confirm the brand, whether the name is stable, and whether the brand sits in the category the query is about. A stronger article on our own site does not repair those three.

The mechanism, with the study-by-study multipliers, is Entity Mass: How Brand Weight Across the Web Shapes. That page is the deeper resource. This essay is how I would use it as a founder: which multiplier I would buy first, which average I would refuse to treat as a universal quota, and what result would send me back to publishing pages.

The split I would put in front of a content plan

Hexagon's analysis of 50,000 AI product recommendations found that 82% of e-commerce brands receive zero mentions across ChatGPT, Perplexity, and Claude. In the same analysis, 3% of brands capture 71% of generative search recommendations. In competitive verticals the study names, SaaS, electronics, and skincare, the top five brands take 83% of recommendations.

Searchless.ai tracked 500 brands across 9,000 queries per brand over 90 days and found that 73% of brands ranking in Google's top three are invisible to AI engines. Those 500 brands averaged $47 million in revenue and $1.2 million a year in marketing spend. Of the 500, 440 were never cited once. I would read that as a separation between ranking and recommendation. A company can win the results page, spend at that level, and still be unnamed in the answer. I would not approve another year of the same marketing mix without a count of independent domains that mention the brand by name.

What I mean by the weight

Entity mass, as these studies use it, is the accumulated weight of cross-domain signals that lets an engine treat the brand as one recommendable thing. In many answers the engine is activating an entity it already represents, rather than discovering a page in the moment. If the brand is not a distinct entity in that representation, the page never gets a chance.

Searchless.ai names three ways a brand stays under the line. Fewer than four independent domains of mention lines up with effectively no AI visibility, while brands on 15 or more domains appear 4 to 7 times more often. The correlation they report between mention density across six or more independent domains and citation frequency is r = 0.78. Inconsistent identity splits the weight: a short name, a legal name, a product name, and a social handle can be four partial entities, and each variation dilutes the record instead of adding to it. Context is the third failure. Being known is not the same as being filed under the query a buyer uses. Co-occurrence in the training record decides which category questions activate the brand. A famous name in the wrong category is absent for the questions that pay.

I would diagnose our record against those three before I argued about which article to write. Domain count, name variants, and category co-occurrence are checkable. "We need more content" is not a diagnosis.

The multipliers, in the order I would fund them

Hexagon's 50,000-recommendation study puts multipliers on specific signals. I would fund them in leverage order, and I would keep each multiplier attached to the study so nobody turns it into a guarantee.

Wikipedia and knowledge-graph presence carries a 9.4x multiplier, the highest in that list. A verified Wikipedia entry is third-party consensus that the brand is a distinct entity. On Claude, the companion analysis of 10,000 citations reports a 5.1x citation-frequency increase for brands with a Wikipedia presence. I would pursue notability that can survive Wikipedia's own rules. I would not commission a page we cannot defend.

Review density carries a 6.3x multiplier, defined there as 500 or more reviews across Google, Trustpilot, and niche sites. The point of the volume is sustained presence, not a launch-week spike. Each review platform is another domain saying the brand is real and in market. I would fund review acquisition only where customers already have a reason to write. A purchased burst fails the "sustained" part of the finding.

High-authority media, placements on DA 70+ publications inside 24 months, carries a 5.9x multiplier. Brands with that coverage had a 71% AI visibility rate against 12% without it. A separate analysis put 85.5% of non-paid AI citations on earned media rather than brand-owned content. I would buy the placement that a DA 70+ desk will actually publish, and I would judge it by whether the sentence names us in the category buyers use.

Citing-domain breadth is the composite. AI-visible brands in the Hexagon set average 47 unique citing domains. Invisible brands average 6. That gap, about eightfold, is the strongest composite predictor in the dataset. I would treat 47 and 6 as markers from this study, not as a quota every category must hit. Searchless.ai's 15-domain and 4-domain markers are a different cut of a related question. I would report our number beside both, and I would close the gap we can see, rather than announce that 47 is the universal threshold.

Structured data is the on-site half. It is present on 91% of cited brands' pages and 23% of invisible brands' pages in the Hexagon set. Searchless.ai found engines extracting the first two sentences 73% of the time when they do retrieve a page, which means a soft opening spends the extraction slot. In that study, 96% of invisible brands lack llms.txt and 94% lack FAQ schema. I would fix Organization schema, a stable opening fact, and the files a crawler is documented to read. I would not call that the strategy. It is the part we control after the cross-domain record exists.

Only 11% of brands in the 10,000-citation analysis had optimized for generative discoverability, and that 11% captured 38% of citations. I would rather be in the smaller group with a coherent record than in the larger group with a heavier publishing calendar.

The failure mode on the famous side of the line

Recognition and accurate recommendation diverge. An arXiv study of 100 entities across 1,400 probe runs found the largest, most familiar brands producing fabricated citations 52.69% of the time, against 37.87% for smaller Tier 3 brands. The gap is 14.82 percentage points, with p = 1.67e-11. Regulatory-framed queries pushed fabrication to 56.77%. The authors describe the mechanism as ghost cartography: a familiar brand sits in a denser region of the model, and the model completes details from neighboring representations with confidence.

I would treat that as a quality constraint on the spend. Enough cross-domain presence to be recognized, and the presence has to be structured, consistent, and factually grounded in the queries where we want to be named. Volume of awareness without a checkable fact is how a well-known company earns invented details. If we are early, the Hexagon concentration says the risk is absence. If we are already widely known, the arXiv result says the risk is a confident error. The repair in the second case is a correct, repeated fact, not more fame.

How I would measure it this month

I would count unique domains that mention the brand by name, using the referring-domain tools we already pay for, and I would put 47 and 6 on the page as Hexagon's markers and 15 as Searchless.ai's marker for a denser footprint. I would not invent a third threshold.

I would search the brand on Google and record whether a Knowledge Panel appears. Wikipedia's 9.4x multiplier is the strongest single signal in the Hexagon list, and a missing panel is a missing confirmation that the public graph has resolved us.

I would run category queries, not the brand name, in ChatGPT, Perplexity, and Google's AI Mode. "Best" and "top" questions in the words a buyer uses. Absence on those queries is the operational definition of being under the line for the revenue we want.

I would list every public variant of the name and collapse them to one canonical form on the surfaces we control, then correct the ones we can ask a publisher to correct. I would run the homepage through Google's Rich Results Test and require Organization schema with sameAs links to the profiles that actually exist, Wikipedia, Wikidata, LinkedIn, and Crunchbase where we have them. A sameAs link to a profile we do not have is a fiction. I would not add it.

The frame I use for why a placement feeds this record, rather than sitting in a clip file, is Machine Relations. The term I would put next to the domain count, so the team does not confuse it with a backlink total, is the entity chain.

I would write the diagnostic onto one page a finance partner can read without the studies open. Our citing-domain count, beside 6 and 47 from Hexagon and beside 4 and 15 from Searchless.ai. Whether a Knowledge Panel exists. Whether Wikipedia is the thing producing the 9.4x, or whether we are hoping a page we cannot defend will. Review count against the 500-review definition of the 6.3x, not against a vibe. DA 70+ placements inside 24 months, next to the 71% versus 12% visibility rates. The first two sentences of the homepage, because extraction hits them 73% of the time. Schema present or absent, remembering that 91% of cited brands' pages had it and 23% of invisible brands' pages had it, and that 96% and 94% are the llms.txt and FAQ-schema gaps among invisible brands in that study. None of those lines authorizes a new slogan. Each of them authorizes a specific repair.

The decision

I would freeze net-new thought-leadership pages for one cycle and spend that cycle on the five checks above. The first invoice I would approve is the one that adds an independent citing domain in the category buyers already use, or that removes a name variant. The first invoice I would decline is another article on our domain that restates a claim no third party has made.

Scope: this follows the three studies named here, Hexagon's 50,000 recommendations and 10,000 citations, Searchless.ai's 500 brands, and the arXiv probe of 100 entities. It applies to a company that is absent from category answers or that is present and being described incorrectly. It does not set 47 domains as a contract, and it does not say a Wikipedia page can be bought as a service.

What would change my mind: I would go back to publishing owned pages first if a later study of similar scale showed owned pages cited as often as the earned and cross-domain signals Hexagon and the 85.5% earned-media figure describe. I would also drop the "four domains is effectively zero" operating line if Searchless.ai or a replication revised that marker with a disclosed method. I would not drop it because a single prompt returned our name once. One prompt is not a replication of 9,000 queries over 90 days.

FAQ

What would I count as entity mass in a monthly review?

I would count independent citing domains, name variants still in public use, knowledge-graph presence, review volume across Google, Trustpilot, and niche sites, DA 70+ coverage inside 24 months, and whether Organization schema points at real profiles. Hexagon's multipliers are 9.4x for Wikipedia, 6.3x for the review density it defined, 5.9x for that media bar, and an eightfold gap between 47 and 6 citing domains. Searchless.ai's companion markers are under four domains versus 15 or more, with r = 0.78 across six or more domains. I would show our counts beside those markers and label the markers as study-specific.

Does a top Google ranking substitute for this?

In the Searchless.ai panel it did not. Seventy-three percent of brands in Google's top three were invisible to the AI engines, inside a set of 500 brands, 9,000 queries each, over 90 days, with average revenue of $47 million and average marketing spend of $1.2 million. Four hundred forty of the 500 were never cited. I would keep investing in the results page for the traffic it still sends. I would not tell the board that the ranking is the recommendation.

How should I read the 47-domain average?

As Hexagon's average for brands that were visible in that dataset, against an average of 6 for brands that were not. I would also keep Searchless.ai's 15-domain marker on the same page. Category concentration differs, so I would not write 47 into a goal sheet as a universal pass line. I would write the gap between our current count and the visible brands in our own category, and I would close that.

What if we are already well known and the answers are wrong?

I would spend on consistent, checkable facts rather than on more awareness. The arXiv probe of 100 entities and 1,400 runs found fabricated citations at 52.69% for the most familiar brands and 37.87% for smaller Tier 3 brands, a 14.82 point gap at p = 1.67e-11, rising to 56.77% on regulatory-framed queries. Fame made the completion more confident. It did not make it more accurate. The 11% of brands that had done the discoverability work and captured 38% of citations in the other Hexagon cut are the comparison I would use: a smaller, cleaner record beat a larger, noisier one.