Tag Archives: artificial intelligence

Own the Dictionary: Build the Glossary Nobody Else Bothered to Write

Leave a reply

Everybody wants to write the definitive guide. Almost nobody wants to write the definition. That gap is the cheapest authority play left on the internet, and it is sitting there unclaimed like a live straddle nobody noticed.

  1. Definitions are the most retrievable prose there is. Roughly 70% of Google’s featured snippets are paragraph snippets – the definition box – and Semrush pegs the winning ones at 40 to 60 words. Sixty words. You have written longer text messages declining a dinner invitation.
  2. Nobody owns the vocabulary of your niche. There is no rights-holder for “polarized range” or “minimum defense frequency.” The dictionary of your industry is unsurveyed land, and the deed goes to whoever bothers to walk it first.
  3. Forty terms, not four. Ahrefs found a single article holding 4,658 featured snippets across the keyword universe. Forty entries is forty front doors, each catching phrasings you would never have thought to target on purpose.
  4. The long tail is where definitions live. Backlinko found 54% of featured snippets come from searches doing under 50 queries a month. Semrush found 55.5% of ten-word queries trigger a snippet versus 4.3% of one-word queries. Nobody wins “poker.” Anybody can win “what is a donk bet and why is it called that.”
  5. Write the disagreement – that is the entire moat. Any intern can type “a donk bet is a lead into the previous street’s aggressor.” The version that gets quoted for a decade adds that purists restrict it to the flop, that it was coined as an insult, and that solvers have since made it standard – meaning half the internet is using a slur to describe correct play.
  6. Three more you can steal tonight. “GTO” strictly means a Nash equilibrium in a heads-up zero-sum game, and its guarantees quietly evaporate three-handed – yet it gets used daily to mean “whatever the solver said.” Sklansky defined a semi-bluff as a bet that is probably not best but can improve; today people call king-high a semi-bluff. And a “merged” range means value-plus-marginal to one camp and close to the opposite to another. Say so. Cite both. Now you are the referee.
  7. Front-load or forfeit. About 44.2% of AI citations are pulled from the first 30% of a page. Term, one clean sentence, then the nuance. Bury the answer and you are bluffing into a machine that folds instantly and never pays you off.
  8. Mark it up. Schema-tagged pages are cited roughly 2.3x more often in AI Overviews. Definition schema takes fifteen minutes. That is the best price-to-equity ratio on this page.
  9. The door is open right now. Ahrefs’ March 2026 data shows only 38% of AI Overview citations come from top-10 organic results, down from 76% a year earlier. You no longer have to outrank Wikipedia. You have to out-define it – and Wikipedia is not allowed to have an opinion about which definition is better. You are.

Guides get skimmed and forgotten. A glossary gets bookmarked, linked, argued with, and quoted – by people and by models – for years. Write forty definitions, note who disagrees and why, and stop playing the hand. Become the house.

Drop the AI Prefix

Leave a reply

Why “AI-powered” stopped being a magnet in 2026 and started working like a coupon.

1. The label now costs you money. Researchers at Washington State and Temple put the same products in front of 1,000 people and changed exactly one thing: whether the description said “artificial intelligence.” Across eight product and service categories, purchase intent fell every single time. It fell hardest where the stakes were highest, a shrug for a television, a wince for a car or a medical diagnostic. Parks Associates asked roughly 4,000 Americans a similar question: 18% said AI made them more likely to buy, 24% said less likely, and 58% said it changed nothing. You are spending the best three seconds of your homepage on a word that repels more people than it attracts and bores the majority.

2. We have seen this movie. It came out in 1998. Cooper, Dimitrov and Rau published the autopsy in the Journal of Finance: companies that simply bolted “dot.com” onto their names earned about 74% abnormal returns in the ten days after the announcement. No new product. No new customers. A new suffix. Books-A-Million announced a redesigned website the day before Thanksgiving 1998 and watched its stock run from $3.06 to an intraday high of $47. Two weeks later it was $10. By 2000 it was back around $3.60. Same bookstore the entire time.

3. The real tell is the deletion trade. Here is the part everyone forgets. When the bubble burst, the same researchers tracked firms that stripped “dot.com” back out of their names and found cumulative abnormal returns of roughly 64% over the sixty days around the announcement. The word paid 74% going on and 64% coming off. A label that is valuable in both directions is not an asset. It is a costume, and eventually the audience notices the zipper.

4. The prefix is now a legal exposure, not a marketing one. The SEC has a name for it, “AI washing,” and a growing docket. Enforcement opened in March 2024 against two advisory firms and reached its first public company in early 2025: Presto Automation, whose drive-thru voice AI turned out to belong to a third party and to lean heavily on humans. The founder of Nate Inc. was charged over more than $42 million raised on claims of an AI shopping app. Six named cases since March 2024, over $44 million in alleged fraud, and securities class actions over AI claims roughly doubled from 2023 to 2024. “Everybody says it” has never been a legal defense.

5. The buyer already assumes it, and has already been burned. In 2019 MMC Ventures reviewed 2,830 European “AI startups” and found no material evidence of AI in about 40% of them, even as the AI label pulled 15% to 50% more funding than plain software. That arbitrage is closed. MIT’s 2025 enterprise study weighed $30 to $40 billion in spending against a blunt finding: 95% of generative AI pilots produced no measurable impact on profit and loss. Your prospect has a dead pilot in a drawer. “AI-powered” no longer excites her. It starts an interrogation.

6. So name the job you finish. “AI-powered claims processing” describes your plumbing. “We close claims in four hours instead of nine days, and you pay only for closed claims” describes her Tuesday. Run the ninety-second audit: delete every instance of “AI” from your homepage. If what remains still names a verb, a number and a deadline, you had positioning. If the page goes blank, you had a costume.

The bottom line. Nobody bought a book in 2004 because the store was “internet-enabled.” They bought it because it showed up in two days. The prefix was never the product, and in 2026 it is not even the hook. Tell me what you finish, how fast, and what it costs. I will assume the robots.

500 fictional venture capitalists at an open bar

Leave a reply

This morning I put 500 fictional venture capitalists in a building.

I broke them into 100 rooms, five people each, and had them argue about a client’s funding strategy. Standard stuff – I run persona work like this on most engagements, whether the question is fundraising, positioning, or go-to-market.

Then I added an open bar.

Tipsy VCs, it turns out, are more useful than sober ones. The sober rooms produced the answers you’d expect: raise a priced seed, here’s your comp table, here’s your dilution. The bar rooms started saying things like “You’re running a restaurant and you’ve never asked what the food costs.” Blunter. Less hedged. Occasionally wrong in interesting ways.

To be clear about what this is: I’m not simulating market feedback, and no fictional VC is going to tell me what a real one will do. What this does is widen the range of angles I have to consider before I am on the call. The unhedged version of an objection is easier to prepare for than the polite one.

The client got better funding options and a cleaner business structure out of it. I credit the bar.

Answer in Spanish: The Second Court Nobody Is Playing

Leave a reply

Your English authority stops at the net. In U.S. insurance distribution, the other court is empty.

1. Clay is not grass. Rafael Nadal won Roland Garros fourteen times. That bought him zero free points at Wimbledon, where he won twice. Same forehand, different surface, different bounce, different everything. Language models work the same way: each language is a partially separate knowledge space. Ranking #1 in English does not seed you in the Spanish draw. Nobody carries a ranking across the net for free. Surfaces are learned.

2. The 97-to-19 problem. Researchers taught a model new facts in English. Quizzed in English, it recalled them 97% of the time. Quizzed on the identical facts in another language, accuracy fell to 19%. Same model, same facts, one language apart. A second study estimates perfect cross-language sharing could lift accuracy up to 150% – a polite academic way of saying it does not happen today.

3. Untranslated means invisible. If your annuity explainer and licensing FAQ exist only in English, a Spanish query does not rank you low. It does not see you. You are not losing the point; you never entered the tournament. English is 49.7% of identifiable web content, Spanish just 6.0% (W3Techs, June 2026).

4. The crowd is already in the stadium. 44.9 million U.S. residents speak Spanish at home (2024 ACS) – 1 in 7 people age five and up. That group grew 21.3% from 2010 to 2024 while the population grew 11.2%. And 41% say they speak English less than “very well.” They are asking chatbots about deductibles, beneficiaries and IUL right now, in Spanish.

5. The number that should sting. Hispanic life insurance ownership fell from 51% in 2021 to 40% in 2025, the lowest of any group LIMRA tracks. Roughly 20 million Hispanic adults say they are underinsured, and 72% overestimate what term costs. That is not a demand problem – the demand is already there. It is an answer-supply problem, and the answers are sitting in the wrong language.

6. Do not practice against a ball machine. Auto-translating your English pages is hitting against a ball machine and calling it a match. “Rider,” “surrender charge,” “face amount” – rendered literally, they land as nonsense that no Spanish speaker and no model treats as authoritative. Write it natively, in the vocabulary agents actually use in Miami and Houston.

7. Serve first. Carlos Alcaraz did not inherit a grass game. He built one on purpose, then won Wimbledon in 2023 and 2024. The Spanish-language retrieval surface in insurance is wide open and lightly defended. Every month you leave it untranslated is a free point handed to whoever translates first.

Searching Your Own AI Archive Is a Trip Back to 1997

Leave a reply

You worked on it three days ago. The keywords were specific, unusual, and yours. You type them in. The first result is from two months ago. The second has nothing to do with anything. The third one is what you wanted.

Anyone who used the web before Google knows that feeling. AltaVista held the whole index and ranked it badly: term frequency over intent, no notion of authority, no sense of when. You learned to scroll, to guess synonyms, to accept that the machine had your answer and could not put it on top.

Recency Is a Signal, Not a Tiebreaker

The failure repeats across Claude, ChatGPT, Gemini and Grok, because retrieval runs on semantic similarity alone. But a person searching their own history is not asking what is most alike. They are asking what did I just do. Recency, revisits, thread length, whether the session actually produced something: all ranking signals, all mostly unused. It is the absence of PageRank over again. The corpus is excellent; the ordering is naive.

Citizen Kane opens in a warehouse of a dead man’s possessions, catalogued and worthless. The one object that explained him sits there, indistinguishable from the crates around it. That is an index without ranking.

The Irony Is Exquisite

The people who solved this are alive and employed. Twenty-five years of work on freshness decay, click models, query intent and personalization sits at Google, at Microsoft, and in every ranking team they trained. The AI labs rebuilt retrieval from first principles and faithfully reproduced 1998.

The fix is not a bigger model. It is an information retrieval hire.

Search companies helping AI companies with search – that is the hoot. Welles had never directed a feature when he made Kane. He also had Gregg Toland, who had shot fifty of them and knew exactly where to put the camera. Vision needed craft standing next to it.

What Good Looks Like

Time-aware ranking on by default. Filters that survive contact with a real question. A working answer to “the last thing I did on X.” And one line telling you why a result matched, so you can correct the query instead of guessing at it. None of this is research. It is product work that was finished a generation ago and simply never carried across.

Kane ends with reporters combing an archive for a single word, failing, while the answer burns in front of them. Our archives are smaller, better indexed, and still ask us to guess. Rosebud was in the room the whole time.

Publish the Questionnaire

Leave a reply

Why your security review belongs on a public URL instead of a locked drawer

Every enterprise buyer sends you the same security, data and model-governance review. Same 200-odd questions, different letterhead. You answer it in week five, under duress, at 11 p.m., by copying last month’s answers and hoping legal doesn’t notice. There is a better plan: answer it once, publicly, as a page anyone can read before the first call.

1. The meeting you are not invited to. Gartner’s B2B buying research is brutal on this point: buyers spend roughly 17% of their total purchase time meeting with all potential suppliers combined, and only 5-6% with any single sales rep. The buying group runs 6-10 stakeholders. Your questionnaire answers attend meetings you will never see. Send them dressed for it.

2. Stop building the Cone of Silence. On Get Smart, Max and the Chief would lower a plexiglass dome over their heads to discuss classified matters, then discover neither could hear a syllable and shout the secret across the room anyway. A SOC 2 report locked behind an NDA and a lead-capture form is the Cone: elaborate, official-looking, and a net loss of information for everybody in it.

3. Answer the test everyone is already grading. The Cloud Security Alliance’s CAIQ v4 runs 261 questions across 17 domains. The Shared Assessments SIG is longer. These are published standards, not state secrets — so post your answers mapped to SOC 2 Type II, ISO 27001, encryption at rest and in transit, subprocessor list, and put a date on every line.

4. Model governance is the new page three, and almost nobody has it. ISO/IEC 42001 landed in December 2023, NIST’s AI Risk Management Framework in January 2023, and the EU AI Act’s obligations for general-purpose models took effect 2 August 2025. Buyers now ask: do you train on our data, what is the retention window, who are your model subprocessors, where exactly is the human in the loop. “We take security seriously” answers none of them.

5. “Would you believe…?” is not a control framework. Max’s signature move was the incremental climbdown — two hundred police cars, then two squad cars, then a Boy Scout with a slingshot. Vague claims deflate the same way under diligence. Numbers do not: 99.9% uptime, 24-hour breach notification, 30-day deletion SLA, AES-256 at rest. Publish the figure you can defend on the record, the first time.

6. Feed Hymie. Hymie the robot was brilliant and utterly literal — tell him to kill the light and he draws his sidearm. Retrieval systems are Hymie. They cannot infer your posture from a brochure adjective. Give them question-and-answer pairs in plain HTML, named standards, real dates, no gated PDF. Gartner projects traditional search volume falls 25% by 2026 as buyers shift to AI answers, and the fact-dense page is the one that gets quoted back.

7. Missed it by that much. Deals rarely die at the demo. They die in week six at security review, when somebody in a room you are not in cannot find an answer and defaults to no. Publishing converts a six-week interrogation into a pre-read, and the call now opens with “we’ve read it, two follow-ups” instead of “send us your SIG.” That is the whole return.

The bottom line: Agent 13 filed excellent intelligence from inside a trash can. That was the problem. Nobody could find him.

Maintain, Do Not Publish

Leave a reply

Every content meeting ends the same way. Somebody asks how many posts we’re shipping. Nobody asks how many we’re fixing. American Idol learned what that costs in front of 30 million people.

𝗧𝗵𝗲 𝗧𝗵𝗶𝗿𝘁𝗲𝗲𝗻-𝗪𝗲𝗲𝗸 𝗦𝗵𝗲𝗹𝗳 𝗟𝗶𝗳𝗲

  1. 𝗧𝗵𝗲 𝘄𝗶𝗻𝗱𝗼𝘄 𝗶𝘀 𝘄𝗲𝗲𝗸𝘀, 𝗻𝗼𝘁 𝘆𝗲𝗮𝗿𝘀. Roughly half the content cited in AI answers is under three months old. Ahrefs analyzed 17 million citations: cited pages run 25.7% fresher than the organic results. ChatGPT is most ruthless: 76.4% of its top-cited pages were updated within 30 days. Your 2023 masterpiece isn’t evergreen. It’s a rerun.
  2. 𝗥𝗲𝗳𝗿𝗲𝘀𝗵𝗶𝗻𝗴 𝗽𝗮𝘆𝘀 𝗯𝗲𝘁𝘁𝗲𝗿 𝘁𝗵𝗮𝗻 𝗽𝘂𝗯𝗹𝗶𝘀𝗵𝗶𝗻𝗴. AirOps tracked 4,000-plus cited pages: 35.2% were updated within three months, 53.4% within six. Refreshed pages average 6 citations versus 3.6 for stale ones, a 67% lift on work you already paid for.

𝗦𝗲𝗮𝘀𝗼𝗻 𝟭𝟮 𝗪𝗮𝘀 𝗮 𝗣𝘂𝗯𝗹𝗶𝘀𝗵𝗶𝗻𝗴 𝗦𝘁𝗿𝗮𝘁𝗲𝗴𝘆

  1. 𝗠𝗼𝗿𝗲 𝘀𝗲𝗮𝘀𝗼𝗻𝘀 𝗶𝘀 𝗻𝗼𝘁 𝗺𝗼𝗿𝗲 𝗿𝗲𝗹𝗲𝘃𝗮𝗻𝗰𝗲. Idol peaked in 2006-07 at 30 million viewers a night. When numbers slipped, Fox did what content teams do: more, louder. The Season 12 finale drew 14.3 million, off 40% in a year and the first ever to miss 20 million. Average audience fell from 23.1 to 13.2 million in two years. Volume wasn’t the fix. Volume was the symptom.
  2. 𝗧𝗵𝗲 𝗼𝗹𝗱 𝗮𝘀𝘀𝗲𝘁 𝗼𝘂𝘁𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗲𝗱 𝘁𝗵𝗲 𝗻𝗲𝘄 𝗼𝗻𝗲𝘀. Carrie Underwood won Season 4 in 2005, became the best-selling Idol winner ever, and returned 20 years later as a judge. Idol didn’t need a new champion; it needed to update the one it had.
  3. 𝗔 𝗻𝗲𝘄 𝗱𝗮𝘁𝗲 𝗶𝘀 𝗻𝗼𝘁 𝗮 𝗿𝗲𝗳𝗿𝗲𝘀𝗵. Fox cut Idol from 50 hours a season to 37 and rotated the panel. The audience left anyway, bottoming at an 8.03 million finale. A fresh timestamp on a 2023 pricing table is a new stage set, same tired song.

𝗡𝗮𝗺𝗲 𝗮𝗻 𝗢𝘄𝗻𝗲𝗿. 𝗧𝗵𝗲𝗻 𝗢𝗽𝗲𝗻 𝗮 𝗖𝗮𝗹𝗲𝗻𝗱𝗮𝗿.

  1. 𝗦𝗼𝗺𝗲𝗯𝗼𝗱𝘆 𝗵𝗮𝘀 𝘁𝗼 𝗯𝗲 𝗦𝗲𝗮𝗰𝗿𝗲𝘀𝘁. Twelve judges have come and gone, from Cowell to Katy Perry. Ryan Seacrest has hosted since 2002. Maintenance without a named owner is a group project, which is to say nobody’s job.
  2. 𝗧𝗵𝗲 𝗰𝗮𝗱𝗲𝗻𝗰𝗲 𝗶𝘀 𝘁𝗵𝗲 𝘀𝘁𝗿𝗮𝘁𝗲𝗴𝘆. Idol opens auditions every August, excited culture or not. Rank your top 20 pages by revenue and refresh five a month; everything gets touched twice a year.
  3. 𝗧𝗵𝗲 𝘂𝗻𝗴𝗹𝗮𝗺𝗼𝗿𝗼𝘂𝘀 𝗲𝗻𝗱𝗶𝗻𝗴 𝗶𝘀 𝘁𝗵𝗲 𝗴𝗼𝗼𝗱 𝗼𝗻𝗲. Idol returned on ABC in 2018 and still runs in 2026, season 24, at 5.9 million viewers a week. Nobody throws a party for updating a pricing page. Do it anyway. The citation goes to whoever showed up this quarter.

Twenty maintained pages beat twenty new ones. You already know which twenty.