How the engines answer
"As of my last update": when an AI engine is not looking you up, it says so
By Arnav Mukherjee, founder of TofuBofu · August 7, 2026
We keep every raw answer our research queries produce, and I was rereading a batch of them last week for an unrelated reason. Twenty five buying questions, put to the six engines we track, collected across July and early August: who are the best managed IT providers in Ohio, who are the best personal injury lawyers in Tampa, who are the best B2B SEO agencies, that sort of thing.
Reading them one after another, the same four words kept arriving at the top of the page. "As of my last update." Not occasionally. I went back and counted, and ChatGPT had opened with that exact phrase in 15 of the 25 answers it gave us.
I had been treating that as boilerplate for about a year. It is not boilerplate. It is the engine telling your buyer, in the first line, that it did not go and look anything up before recommending a vendor, and it is the single most useful signal in the whole answer.
Two engines hedge. Four never do.
Here is the count across all 25 questions. I looked for three things in each answer: an explicit memory caveat such as "as of my last update", an explicit statement about a knowledge cutoff, and a closing instruction telling the reader to go and verify current status before acting.
Fifteen of ChatGPT's 16 were the literal "as of my last update". Claude phrased it differently and, to its credit, more usefully: twice it stated plainly that "my knowledge has a cutoff date", and nine times it closed with a section telling the reader to verify current status, warning that firms merge, close and change hands, and pointing them at Google Reviews, G2, Capterra, Gartner or the local directory. Claude is effectively telling your buyer to go and check a third-party source. Whether you appear in that source is then your problem.
The other four of the six produced 50 answers between them and hedged in none of them. The nearest thing to a date reference anywhere in those 50 runs the opposite way. Perplexity, answering about Calgary providers, wrote that the firms were "verified with real Google ratings as of mid-2026". That is not a caveat about stale memory. That is a claim to freshness, from an engine that had just been out to look.
What a memory answer actually returns
The clearest single illustration in the set is a query we ran on 31 July, asking each engine to name the best AI development companies for building custom AI products. It is a straightforward buying question with an obvious kind of correct answer, namely firms that build custom AI products for clients.
ChatGPT, opening with "as of my last update", returned OpenAI, Google AI, IBM Watson, Microsoft AI, AWS AI, C3.ai, DataRobot, H2O.ai, NVIDIA AI and Palantir. Read that list as a buyer with a budget and a product to build. It is model providers, cloud platforms, a chip maker and enterprise software vendors. It is, almost exactly, a list of the things an AI development firm would buy in order to serve you, rather than firms you would retain to do the building. A couple of them will take on large custom engagements if you are an enterprise, but as an answer to the question asked it is a supplier list wearing a vendor list's clothes.
Claude's answer had the same shape with a partial rescue: the same platform giants at the top, then McKinsey, Deloitte, Accenture and BCG, which at least are firms you can hire, then Scale AI, Hugging Face, Anthropic and Jasper. Jasper writes marketing copy.
The two engines that retrieved answered a recognisably different question. Gemini returned LeewayHertz, Deeper Insights, Addepto, InData Labs, Neoteric, ScienceSoft, Pixelplex, QuantumBlack and IBM Consulting. Bing Copilot returned Codiant, Biz4Group, TatvaSoft, SmartOSC, DICEUS, LeewayHertz, Markovate and InData Labs, before tacking the usual platform names on at the end. These are mid-market and boutique development firms with sales teams and a proposal process.
The overlap is the part I would put on a slide. ChatGPT and Gemini, asked the same question on the same day, shared not one company name. Claude and Gemini also shared none. The two retrieving engines, meanwhile, independently landed on two of the same firms. The engines are not disagreeing about who the best vendor is. They are answering two different questions, and only one of them is the question the buyer asked.
The same buying question, two different machines behind it
The caveat predicts whether the names can be checked
A fair objection at this point is that a hedge is just politeness, and the recommendations underneath it might be perfectly good. We have measured that separately, and they are not.
In our 15-region study of managed IT providers, every company name the engines produced was resolved to a real website and checked against two independent references: the region's own Clutch directory and Google's results for the same question. Perplexity's names could be corroborated 39 percent of the time, 101 out of 257. Gemini managed 18 percent. ChatGPT managed 7 percent, and Claude 2 percent, which is 2 names out of 129.
That ordering is the same ordering as the caveat count, inverted, and it is not a coincidence or a comment on how good these models are at anything else. An engine that retrieves can only name companies it just found evidence for. An engine composing from memory has no such constraint, so it produces fluent, locally specific, largely uncheckable shortlists. The caveat is the model being honest about the method. The corroboration rate is what that method costs.
Find out which engines can actually see you
Run a free scan on your own buying question and read all six engines separately, including the ones answering from memory.
Get your free auditAn important limit on what I am claiming
These answers come from calling the models through their APIs, which is how any monitoring tool collects at volume, and in that mode two of the six have no web access attached. That is the mechanism, not a trick: I am not claiming the ChatGPT app behaves this way. It often does not. Ask the consumer app the same question with browsing available and it will frequently go and retrieve, and then the caveat disappears and the answer improves. We wrote about that gap separately in API versus chat interface, and it is worth understanding before you read any vendor's engine coverage claims, including ours.
What survives that caveat is the thing that matters to you. Both modes reach real buyers. Some of your buyers will ask a model that retrieves, and some will ask one that composes, and a fair number will ask one that decides for itself which to do on that particular question. You do not control which one they get. What you can do is stop reading the two as one number, because the work that wins them is not the same work.
This is also why I keep saying SEO is the floor rather than the whole game. Being crawlable, readable and present in the sources an engine can reach is the precondition for a retrieved answer to name you. It is necessary, and on this evidence it is nowhere near sufficient, because the retrieved answer is decided on corroboration rather than on rank.
What to do with the tell
Four things, roughly in this order.
Read the answer's own language before you read the names
Cutoff phrasing, a verify-before-acting closing section, or the absence of any citation tells you the answer was composed. Citations, current detail such as live ratings, and named sources tell you it was retrieved. This takes five seconds and tells you which of the two competitions you are looking at, which is more than the shortlist itself tells you.
Do not spend this quarter trying to get into a memory answer
There is no submission form for a training corpus. Nothing you publish this month can appear in an answer composed from data assembled before it. Effort aimed there has no feedback loop, which is exactly the kind of work that quietly consumes a year.
Compete where retrieval decides, because that is where you can move
A retrieved answer is built from material an engine can find and read today. Directory listings that agree with each other, a site that states plainly what you do and who for, third-party roundups and reviews, and pages that answer the buying question directly. Breadth of corroboration is the lever with an actual handle on it.
Track the two separately, forever
If you average a retrieved engine that names you with a composed engine that never will, you get a number that moves for reasons you cannot act on. Keep them apart. A rising score on the retrieval surfaces while the memory surfaces sit flat is not a disappointing result, it is the expected shape of progress, and you should be able to see it.
None of this is a reason to shrug at the channel. G2's 2026 buyer research found 51 percent of B2B buyers now begin vendor research on an AI chatbot, up from 29 percent, and 69 percent had changed their vendor choice based on what an AI told them. The buyer who gets the composed answer does not see the caveat as a caveat. They see ten confident names, and they start there. You just have the advantage of knowing which ten those were never going to be.
Frequently asked questions
What does it mean when ChatGPT says 'as of my last update'?
It means the model is composing the answer from what it absorbed during training rather than from anything it fetched while answering. The phrase is a hedge about the age of that training data. For a vendor recommendation it has a sharper consequence than most people read into it: nothing you have published since the training data was assembled can be in that answer, no matter how good it is, because the model never went and looked. In our stored set of 25 vendor-recommendation queries, ChatGPT opened with that phrase 15 times.
Do all AI engines answer from memory?
No, and the split is close to binary. In our 25 queries the two engines we called as plain model APIs, with no retrieval attached, carried a memory or verify-first caveat in 16 and 11 answers respectively. The engines built on retrieval, Perplexity, Google AI Mode and Bing Copilot, plus Gemini in this run, carried one between them, and that one was a freshness claim rather than a caveat: Perplexity said its ratings were verified as of mid-2026. Retrieval-backed surfaces read the live web before answering, so they have no reason to hedge about a cutoff.
Can I get into an answer that comes from trained memory?
Not on the timescale you can plan around. What a model absorbed in training is fixed until the next model, and you have no way to submit to it, appeal to it or buy into it. What you can influence is the volume and consistency of independent material about your company that exists before the next training pass, which is the same work that gets you into retrieved answers. So the honest sequence is to compete where retrieval decides, and treat any improvement in memory answers as a slower second-order effect rather than this quarter's goal.
Does answering from memory make an engine less accurate about companies?
In our data it tracks closely. In a 15-region study of managed IT providers we resolved every company the engines named to a real website and checked it against the region's own directory and Google results for the same question. Perplexity, which retrieves, could be corroborated on 39 percent of the names it produced, or 101 of 257. Gemini reached 18 percent, ChatGPT 7 percent and Claude 2 percent, or 2 of 129. This is a difference in method rather than a verdict on model quality, but it means a memory-composed shortlist is far more likely to contain names nobody can verify.
What is the difference between a memory answer and a retrieved answer for my business?
They are different competitions with different entry conditions. A retrieved answer is decided by what an engine can find and read about you right now, which you can act on within weeks by publishing, getting listed and getting corroborated. A memory answer is decided by how prominent your name was in a training corpus assembled some time ago, which you cannot act on directly at all. Treating them as one channel is what makes AI visibility feel arbitrary. Separating them tells you exactly which effort has a feedback loop attached.
Why did ChatGPT recommend OpenAI and NVIDIA when I asked for AI development companies?
Because a model answering from memory reaches for the most strongly associated names in a category, and in AI the strongest associations are the platforms, not the service providers. Asked in July 2026 to name the best AI development companies for building custom AI products, ChatGPT returned OpenAI, Google AI, IBM Watson, Microsoft AI, AWS AI, C3.ai, DataRobot, H2O.ai, NVIDIA AI and Palantir. Those are model providers, cloud platforms, a chip maker and enterprise software vendors, which reads as a list of the things a development firm buys in order to serve you rather than firms you would retain to build the product. The two retrieving engines answered the same question with actual development firms.
How do I check which kind of answer my buyers are getting?
Ask the buying question yourself on each surface and read the answer's own language. Cutoff language, 'as of my last update', 'my knowledge has a cutoff date' or a closing instruction to go and verify current status, tells you the answer was composed rather than retrieved. Citations, live detail such as current ratings, and named sources tell you it was retrieved. Do this per engine, because the same company can be present in the retrieved answers and absent from the composed ones on the same day, and a blended visibility score will hide exactly that split.
Sources and further reading
- TofuBofu, AI visibility for managed IT providers, 2026: the 15-region study behind the per-engine corroboration rates, including method and per-region data.
- TofuBofu Research: the regional and category studies the 25 stored answers were collected for.
- Lewis et al., Retrieval-Augmented Generation (arXiv:2005.11401): the original result that letting a model fetch documents before answering produces more factual, more specific output than generating from parameters alone.
- G2 2026 B2B buyer research: 51 percent begin vendor research on an AI chatbot, up from 29 percent, and 69 percent switched vendor based on AI.