ChatGPT and Google AI Are Describing the Company You Used to Be: The Refresh Clocks Behind Every AI Answer

The Portrait Is Late, Not Wrong

Entity lag is the gap between what a company is today and the version of it an AI engine describes. It is not a defect, and it is not random. The model serving Google’s AI Mode documents a March 2026 knowledge cutoff, with some domains held at January 2025. The snapshot behind ChatGPT stops at August 31, 2025. Every answer drawn from that memory describes the company those dates remember.

The audience meeting that remembered company is not small. Google’s Q2 2026 earnings remarks put AI Mode past 1 billion monthly active users, with AI features in Search sending billions of clicks to websites every week. OpenAI reports more than 1 billion weekly active users for ChatGPT. Whatever these surfaces believe about a company is the version most buyers now meet first.

This journal has already mapped where an answer comes from: the two-path split between memory baked into the model at training and a live fetch at question time is covered in parametric memory versus live retrieval. This article is about the dimension that piece left open: time. Each path runs on a documented clock, and the clocks are slower than any change a company makes.

The argument here is that the lag is measurable, then compressible. The engines publish their cutoffs, their refresh mechanics, their correction channels. Read together, those documents form a matrix of refresh vectors with known speeds. The DSF Refresh Clock Matrix, developed later in this article, turns that reading into an audit an executive can run on their own company this week.

The symptom rarely announces itself. It surfaces as a prospect who arrives convinced the company still lacks the capability it shipped last spring, or a candidate who declines because the assistant summarised a culture three leadership changes old. No dashboard measures descriptions. Analytics counts visits, rankings count positions, and the words the machine uses to introduce the business go entirely unwatched.

The Cutoff Ledger: Where Each Engine’s Memory Stops

Start with Google, because Google’s AI surfaces are the largest. The official model card for Gemini 3.8 Flash, the model Google lists as distributed in the Gemini app and Google AI Mode, states a knowledge cutoff of March 2026, with knowledge in some domains limited to January 2025, in line with the wider Gemini 3 family. That January 2025 line sits twenty months behind this article’s publication date.

OpenAI publishes the same ledger for its side. The gpt-5.2-chat-latest page documents the GPT-5.2 snapshot used in ChatGPT with an August 31, 2025 knowledge cutoff, while the model registry lists April 30, 2026 for the flagship GPT-6 Astra. None of this is hidden. The dates sit on public specification pages that almost no buyer, and few sellers, ever read.

The cutoff is not negligence. Training a frontier model is a discrete industrial event: a corpus is assembled, frozen, then burned into weights over months of computation. Nothing about the process resembles a feed. Whatever the corpus held on freeze day is what the model knows, until the vendor pays for another run. The portrait updates the way a printed encyclopedia did, edition by edition, never page by page.

Now put a company’s last two years against those horizons. A repricing shipped in March, a leadership change in June, an acquisition closed in January, a product line retired last autumn: each falls inside at least one model’s blind window, several fall inside all of them. The machine is not misinformed about these events. It is structurally unaware they occurred.

The Sticker Date Is the Optimistic Number

The published cutoff is the best case, and research says the truth runs older. The Dated Data study, presented at COLM 2024, separated reported cutoffs from effective ones by dating what the training corpora actually contain. In one 2023 snapshot, over 80 percent of the Wikipedia documents were earlier versions of their pages. The dump was labeled 2023. Its knowledge largely was not.

The mechanism behind that gap is mundane. Training corpora are assembled from crawls of crawls, archives of archives, so the copy of a page that survives into the corpus is often an old capture rather than the live version. Deduplication then favours the copy that exists in the most places, which is the oldest one. The freeze date describes when the archive was sealed, not how old the things inside it were.

Staleness has also been measured directly at the answer layer. The DyKnow benchmark, published at EMNLP 2024 Findings, asked models about time-sensitive real-world facts and dated the replies: 13 percent of GPT-4’s answers were outdated, rising to 35 percent for the GPT-3.5 ChatGPT of that era. Those are older models, cited here as proof the failure mode is real and counted, not as current rates.

For a company, the implication inverts the usual freshness instinct. The newest fact about the business, the one marketing is proudest of, is precisely the fact least likely to exist in any model’s memory. The older the claim, the more training runs have absorbed it, the more corroborating copies exist, the harder it holds on. Memory favours the past by construction.

When the Engine Looks and When It Remembers

If retrieval fired on every question, cutoffs would not matter. It does not. Google’s grounding documentation states that the model analyzes the prompt and determines whether a Google Search would improve the answer. The fetch is a judgment call the engine makes per query, made precisely to reach beyond the knowledge cutoff, which is Google’s own framing of what grounding exists to fix.

ChatGPT works the same way. OpenAI’s search documentation says the assistant may search the web automatically when a query benefits from current information. The operative word is current: a question about scores, prices, or news advertises its own freshness need. A question about what a company does, or which vendors lead a category, reads as settled knowledge. Settled questions get remembered answers.

Run the distinction on real queries. Ask either engine for a share price, a weather forecast, or last night’s score, and the fetch fires every time, because the question itself names the present. Ask what a specific consultancy does, who owns it, or how it compares with a rival, and the phrasing carries no time signal at all. Identity questions are grammatically timeless, so the engine treats the answer as settled and reaches for memory first.

That is the trap for company facts: they look timeless to the engine while being the facts a business changes most. And a fetch, when it does fire, is no guarantee of the present, because retrieval has its own bias toward established pages. This journal covered that selection effect in stale-source bias: even the live path can return the old you.

The probabilistic path has a measurement implication. The same question asked twice can travel different routes, one reply grounded with cited sources, the next served from memory with none. An audit that records only what the engines said is half an audit; record whether each answer carried citations, because a cited reply dates from the last crawl while an uncited one dates from the training freeze.

The Buyer Meets the Old You

The commercial weight of the lag depends on who consults the portrait, and the answer is now: the start of nearly every serious purchase. Forrester’s State of Business Buying 2026 reports that generative AI searches have become the starting point of the B2B buying journey, feeding a process that involves thirteen internal stakeholders plus nine external influencers before a deal closes. The first brief those twenty-two people share is the machine’s.

Gartner’s numbers close the loop. In its survey of 646 B2B buyers, 45 percent used AI during a recent purchase, while 67 percent said they prefer a buying experience with no sales rep at all. Read those together: the buyer consults the machine, then avoids the one conversation in which a human could correct what the machine remembered wrong.

The rep-free preference removes the correction layer sales once provided. When two thirds of buyers want no conversation, the stale claim is never challenged by a human who knows better; it simply shapes which vendors reach the shortlist. Prevention is the only defence left, which is why the portrait has to be corrected at the source before the buying committee ever convenes.

The stakeholder count is what turns one stale answer into an institutional belief. The first researcher pastes the machine’s summary into a vendor-comparison document, the document circulates to the other twenty-one people, and the stale portrait becomes the deck everyone argues from. Nobody re-asks the question. A description that was wrong for one reader in minute one is wrong for the whole committee by week three.

So the lag window is not a cosmetic embarrassment. It is a period in which demand is routed by a description of the company that the company has already retired. Ask both engines the ten questions a buyer would ask, and read what comes back with a date in mind. If what comes back is the business from two pricing models ago, that discovery is what Answer Engine Optimization (AEO) exists to repair.

The Four Clocks You Can Actually Wind

One clock is out of reach: nobody schedules a model rebuild for a customer. The other three take input, and the fastest belongs to Google’s Knowledge Graph. Google’s knowledge panel documentation describes panels as automatically generated and automatically updated as the web changes, with a claim channel for the entity itself. Feedback from the verified entity is reviewed within a few days. Days, against a twenty-month parametric floor.

The second windable clock is the page layer that retrieval reads. Keeping the canonical entity pages current is necessary but not sufficient, because discovery has latency: the IndexNow protocol exists precisely because unpushed changes can take days or even weeks to be found. Publishing Organization structured data turns the homepage into a machine-readable identity record rather than prose a parser must guess at.

The third clock is access. OpenAI’s publishers FAQ is blunt: appearing in ChatGPT summaries and snippets requires allowing the OAI-SearchBot crawler, a one-line robots.txt decision many security teams made without the revenue team in the room. Beyond access sits corroboration, making the independent sources both engines retrieve agree with the new facts. The accuracy playbook for that layer lives in controlling what AI says about your business; here it is stage four of five.

Corroboration deserves its own mechanics, because engines weigh agreement. A retrieval pass that finds one fresh homepage against ten directories, profiles, and old articles still describing the previous company reads the ten as the consensus. Winding this clock means walking the outer ring in order of retrievability: the profiles engines actually cite for the category first, the long tail after, until the fresh version is the agreeing majority.

The winding order matters more than the effort. Marketing instinct reaches for new content first, but a fresh thought-leadership piece rides the slowest clocks: it must be crawled, indexed, then judged worth retrieving. The panel correction rides a clock measured in days, touches the card Google shows beside every branded search, and costs an afternoon. Speed per unit of effort, not effort, sets the sequence.

The DSF Refresh Clock Matrix

The DSF Refresh Clock Matrix turns the documentation above into a repeatable audit. The matrix itself crosses the two engines against the four refresh vectors, each cell carrying its documented clock speed and its lever. The five stages run it against one company, and the output of stage two, a single number of months, is the figure that belongs in front of a board.

Stage one asks ChatGPT and Google’s AI surfaces the ten questions a buyer would ask, recording every factual claim verbatim. Stage two dates each claim to the era it was true; the oldest date is the entity lag. Stage three places every stale claim in its matrix cell: a wrong founding-era description rides the weights, a dead product in the panel rides the Knowledge Graph, an old price in a cited answer rides the crawl.

A worked stage two makes the number concrete. A firm that rebranded its services in February 2026 asks both engines the ten questions in September. Google’s AI Mode, on its March 2026 cutoff, may already carry the change; ChatGPT’s August 2025 snapshot cannot. If the oldest claim still being served dates to a pricing model retired in 2024, the entity lag is roughly twenty months, and that is the number the board hears.

“The machine is not wrong about your company. It is late, and late is expensive.”

— Digital Strategy Force, Search Intelligence Division

Stage four winds what turns: claim the panel and correct it as the verified entity, bring the canonical pages current, push the changes rather than waiting to be found, open the crawler gates, then build corroboration outward. Stage five calendars the re-measure against release events, because OpenAI publishes dated model release notes while Google ships dated model cards, and each release resets which era the machine remembers.

One caution belongs in the plan. Winding clocks means publishing genuine changes to the surfaces machines read, never simulating freshness. Re-dating unchanged pages or churning cosmetic edits to look recent is precisely the deceptive-freshness signal search spam systems now police. The audit works because it moves real facts closer to the machine, not because it decorates old facts with new dates.

Ownership decides whether any of this happens. Messaging belongs to marketing, robots.txt belongs to engineering, the panel claim sits with whoever once verified it, so the machine’s portrait of the company belongs to nobody. The matrix is also an assignment sheet: every cell has a lever, every lever needs a named owner, and the entity lag number gives the board a way to hold that owner to a trend.

The asymmetry is the strategy. The clock that cannot be wound is slow, indifferent, owned by the vendors. The clocks that can be wound are fast, cheap, owned by the company. A business that keeps the fast clocks telling the present forces every grounded answer to contradict the stale memory, which is the practical version of the moat this journal described in memory, not content: until the next rebuild absorbs the new you, the evidence layer is the whole game.

FAQ — AI Entity Lag

Next Steps — Run the Lag Audit

▶ Ask ChatGPT plus Google's AI surfaces the ten questions a buyer would ask about the company, then record every factual claim verbatim.

▶ Date-stamp each claim to the era it was true; the oldest date in the set is the entity lag the program must close.

▶ Claim the knowledge panel, correct it as the verified entity, and log the review turnaround.

▶ Update the canonical entity pages, add Organization structured data, allow OAI-SearchBot, and push changes rather than waiting for discovery.

▶ Calendar the re-audit against model release notes, because every release resets which version of the company the machine remembers.

The audit is one afternoon of asking the machines what they believe about the company. Compressing the lag they reveal is a program with a winding order, and Answer Engine Optimization is where that program gets run.