Why Part Of Your AI Authority Takes Years, Not Campaigns & Why It Comes From Other People via @sejournal, @DuaneForrester
Webinar New AI Search & SEO KPIs: 4 Real Signals AI mentions and citations are benchmarks, not decisions. Get 4 traffic-predictive signals drawn from real bot data across hundreds of sites. Rundown Inside AI Max, PMax & Smart Bidding What to measure, which signals to prioritize, and how to guide automation before AI Max becomes the default. Guide Local Google Visibility Guide + Cheat Sheet Track how your business appears across Google Search, Maps, and Gemini. Listings, reviews, and competitor signals all in one view. Webinar New AI Search & SEO KPIs: 4 Real Signals AI mentions and citations are benchmarks, not decisions. Get 4 traffic-predictive signals drawn from real bot data across hundreds of sites. Webinar How Freshpet Earned AI's Trust The GEO playbook behind Freshpet's AI Overview citations, plus checks to run on your own brand. Free, live Aug 20. 🔥[Live 8/12 with Loren Baker] Ecommerce SEO: Own your "brand +promo code" search. What models say about your company comes from how other people have described it over time. Today’s work shows up in the next model generation, not this one. Duane Forrester 13 hours ago ⋅ 13 min read Duane Forrester Founder and CEO at UnboundAnswers.com Bio Follow 63 READS A reader left a comment on the entity mapping article suggesting that the obvious next move was to go win the parametric side. That piece had drawn a line between what a model retrieves at the moment you ask it something and what it already carries in its weights. That comment, I think, accepted the line was there, and then treated one half of it as a work item. The phrasing is everywhere right now. Influence the parametric side. Build parametric authority. Four words, verb first, sounding like something you assign to someone with a deadline. I understand why it gets said that way, and the shorthand is doing something useful. The problem is that the work behind those four words was finished years ago in most cases, done for reasons that had nothing to do with language models, and nobody recorded it as a cost at the time because there was no category to record it against. There is a long habit in this industry of taking something nobody controls and giving it shorthand that sounds controllable. Proxy metrics have served practitioners well for exactly that reason, because they give you a usable number for a system you cannot observe directly. That was never the failure. The failure was always in mistaking having a broad understanding for having tight data. What is different this time is what got compressed. A proxy metric reduces a system nobody can see whole to a single number, and everyone using it understands the number is a stand-in. Influence the parametric side reduces years of work across several departments to a single instruction, and nothing in the phrase admits to standing in for anything. The noun is accurate. The verb is not. The years are not the point. What matters is how many separate parties described the company, and how differently each of them said it. Years are simply how long that usually takes to accumulate. Parametric standing, meaning what a model already says about you before it looks anything up, is the result of that accumulation. Two terms get used interchangeably here, and they are not the same thing, which matters for what follows. Training data is the text that went in. Parametric standing is what survived compression into the weights. The relationship between them is real and directional. That is why the research below can measure one against the other and get consistent answers. It is also lossy. A great deal goes in and does not come back out in usable form, and that gap is the source of most of the confusion here. The research here is unusually direct. Kandpal and colleagues found at ICML that a model’s accuracy on a fact tracks the number of relevant documents it saw during pretraining, and they established that relationship causally rather than only correlationally. They also estimated that models would need scaling by many orders of magnitude before answering competitively about subjects with thin support in the data. Waiting for a bigger model does not fix a thin footprint. Mallen and colleagues reached the same limit from another direction, finding that models handle well-covered entities and struggle badly with the long tail, and that scaling mostly improves recall at the popular end while leaving the tail roughly where it started. Then there is the finding this whole argument rests on. Allen-Zhu and Li showed that knowledge only becomes reliably extractable when it turns up in sufficiently varied phrasing during pretraining. Without that variation, a fact can sit inside the model and still return zero percent accuracy under questioning, present but unusable. Their work runs on a controlled dataset and the recommendation is aimed at engineers building pretraining pipelines, so I would not stretch it into a claim about brands. As a mechanism, though, it explains a great deal: repetition from a single source does not produce what variety from many sources produces. Parametric standing is not earned by publishing more. It is earned when enough independent parties describe a company, in enough different ways, that the description survives compression into the model’s weights. Volume of self-published content does not substitute for variety of independent description, because the mechanism rewards distinct phrasing from separate sources rather than repetition from one. Still, somebody can point at a company founded in 2023 that current models describe accurately, and that deserves a clear answer. It did not beat the mechanism. Enough separate sources described it at once that the accumulation happened fast. The exception runs on the same rule. The concept of “viral” applies here, too, just like in social media. The bottom line here is this: more content alone doesn’t work here. You still need to influence people to speak about your company in the way you want, to have the desired impact. And that takes a long time to accumulate. More content answering questions directly is useful, but still takes time to influence the process. The second reason this cannot be tasked is that there is nothing to task it against. Elazar and colleagues, in a project called What’s In My Big Data, examined 10 corpora used to train popular models. One of them is C4, the Colossal Clean Crawled Corpus, a public training dataset built by filtering a single snapshot of Common Crawl. They found C4 drawing from such a diverse set of domains that even the single most common one accounts for less than five hundredths of one percent of documents. Your own property, however much you publish on it, is a vanishingly small share of that. Common Crawl’s own published statistics add something anyone who has watched crawl behavior will recognize. They note that their domain rankings only partly reflect the importance of those domains, because the crawler respects robots.txt and works hard not to overload servers, with the result that highly ranked domains tend to be underrepresented, and the crawler favors plain HTML over other document types. A separate paper documenting C4, led by Jesse Dodge, found the same divergence, noting that the sites inside the corpus do not represent the most used sites on the internet. Set that against where people actually discuss companies online and the problem compounds. Several of the places where a business accumulates the most description are large, heavily trafficked platforms of exactly the kind a polite crawler treats gently, and much of what sits on them is rendered rather than served as static HTML. This is the part I keep coming back to, but I want to be careful not to overstate it. Language models entered general use about four years ago. The text they were built from is considerably older, in two ways that are documented rather than assumed. The most thoroughly examined corpus available is C4, whose source snapshot was taken in April 2019. The Dodge team sampled a million of its URLs and used the earliest Internet Archive index date as a proxy for when each page was written, estimating that 92% were written between 2011 and 2019. They also noted the date distribution is long-tailed, with a non-trivial amount of material written ten to twenty years before collection. C4 is not what current production models run on, so take it as the best-documented corpus rather than a current one. The pattern still tells you something. More pointed is the work by Cheng and colleagues at Johns Hopkins on effective cutoffs. They found that the date a model reports and the date its knowledge actually concentrates around often differ substantially, for two traceable reasons: New Common Crawl dumps carry some amounts of older material, and deduplication can struggle with semantic and near-duplicate content. The practical consequence is that a model’s picture of a company is probably older than its published cutoff suggests. The description a model carries of a company was deposited years before anyone thought to optimize for it. Mainstream language models have existed for around four years; the corpora underneath them skew substantially older, and research on effective cutoffs indicates that a model’s knowledge concentrates earlier than its published cutoff date implies. Whatever standing a company holds today came from work done when the category did not exist. So that means your structured data, your content, your PR work, your review management from then, had to be best of breed to serve you well today. Here is where it gets awkward for anyone who has run a marketing organization. Almost every function that built this reports to marketing. Public relations, analyst relations, community management, trade and event presence, local press work, crisis communications, review operations. None of it sits outside the remit. It is the ordinary work of the department, year after year. (Depending on your company, IT or Systems teams also have a portion of impact.) What none of those functions ever owned was the output, and this is critical. Public relations earns coverage a journalist writes. Analyst relations earns assessments an analyst forms. Community participation earns descriptions from people with no relationship to the company at all. Crisis response produces press written by parties who are, at that moment, adversarial in some cases. Wikipedia presence depends on editors deciding a company is notable, which is an editorial judgment that has never been for sale. Reviews are the clearest case and the most humbling. A company controls the response. It does not control the review. Influence over what a customer writes depends on marketing, product, service, pricing, and staffing all landing correctly on the same day, and even then the customer writes whatever they want. That record accumulates across thousands of separate nodes, over years, in language nobody at the company chose, attached to a business the executive team is accountable for.
Source: Search Engine Journal
This article has been carefully curated and reformatted for educational and informational purposes. Full credit goes to the original publisher.
📚 Visit more helpful articles on Joab Peters Blog
No comments