Sponsored by
A CRM so smart, it updates itself.
HubSpot's CRM is so smart, it now updates itself. Calls get logged and summarized the moment they end. New leads get researched and a first outreach drafted, automatically. Deals move forward on their own, with next steps flagged before you have to ask.
Even the one repetitive task you've been meaning to fix can now run itself, no code required. Customers using it are already seeing the difference.
This isn't the CRM you have to work hard to figure out. It's the one already working before you log in.
iPrompt
THE AI NEWSLETTER THAT TURNS NEWS INTO ACTION
ISSUE #160 WEDNESDAY · 4 NOVEMBER 2026
THE HOOK
Last month Google’s AI Overview stated, as fact, that about 94% of a website’s first impressions are design-related. The number traces to a 2004 study of 15 British women judging health websites. It counted their comments about rejecting sites, not first impressions. Twenty-two years of copying did the rest. How many numbers in your last deck took the same trip?
OUR ANGLE
A link isn’t a source. Follow it to the study.
Trakkr, which sells AI-search tracking tools, asked ChatGPT, Perplexity, Gemini, Claude and Google’s AI Overviews 200 questions that call for a number. It traced a random sample of 1,285 cited statistics up to five pages deep and checked 150 trails by hand. Its weighted estimate: 40% reached a study that said the same thing, 8% one that said something else, 40% a paywall, a dead page or a source silent on the number, and 11% nowhere at all. The last 1% wouldn’t load.
The figure that bothers me is 82%. That’s how often the cited page did contain the number. So the check most of us run, click the link and spot the figure, usually passes. It just isn’t the check that matters.
ChatGPT led at 57% sourced; the rest sat between 38% and 46%, too close to rank. Agreement didn’t help. Four figures came from all five with no study anywhere, including the old line that winning a customer costs 5 to 25 times more than keeping one. Caveats: one ask per question, via the APIs, by a firm with tools to sell.
My read: an AI-quoted number is a lead, not a source (Trakkr’s phrase, and the right one). Before it goes in a deck or a proposal, follow it until you reach the study, then check who was asked, how many, when and where. The answer box is swallowing the click that used to do that job, and in the US Google now sells ads in its follow-ups. Nobody will check for you.
Prediction: by 30 June 2027, at least one of those five assistants will mark by default whether each cited page is original research. How I’ll score it is in the companion.
GO DEEPER
How to trace an AI statistic: a four-verdict table you can copy into your next deck review, and the check that caught an AI tracer quoting sentences that weren’t on the page.
PROMPT OF THE WEEK
Trace a statistic before you use it
For any number an AI gave you that’s heading into something with your name on it. Use an assistant that can browse, or paste in the text of each page.
I’m about to use a statistic and need to know whether it holds. Don’t defend it. Trace it.
STATISTIC, AS I’LL WRITE IT: [the exact sentence]
WHERE I GOT IT: [the AI answer, and the link it cited if any]
1. Open the cited page and quote the exact sentence containing the number. No citation? Search for the exact claim, and for the original study by name if the answer gives one.
2. If the page is dead, or doesn’t contain the number, try an archived copy (web.archive.org) and the report it names, on the publisher’s own site. Only then stop at CAN’T CHECK.
3. If a page credits another source, open it and repeat, up to five hops. For each hop give the URL, publisher, date and the sentence quoted. If you reach five hops with a lead still open, stop at CAN’T CHECK and show the last link.
4. Stop at original research: a survey, study, dataset or official statistic with a method. Name who ran it, when, how many it measured, where, and the exact question.
5. Compare my sentence with the original on five points: what was measured, the year, the source, who was asked, the place. Flag every mismatch.
6. Give one verdict. SOURCED: the original says what my sentence says. CHANGED: it differs on the number or any of the five points; give me a corrected sentence. CAN’T CHECK: say which wall you hit (paywall, form, dead page, no number, five-hop limit), show the last link, and say what would get past it. NOWHERE: only when the trail ends at pages that credit no study.
7. Don’t fill gaps from memory. A page you couldn’t open is not evidence.
Why it works: step 1 is the check that passed 82% of the time (the number is on the page). The rest push on to the one that passed 40% (a study agrees). Warning: a tracer can invent quotes. Find each on its page.
AI AND ROBOTICS THIS WEEK
A humanoid maker heads for Nasdaq
Agility Robotics, maker of the Digit warehouse humanoid, expects to start trading as AGLT tomorrow, if its merger with the SPAC Churchill Capital XI closes today as planned. Shareholders approved on Monday. Holders of 61.5% of the SPAC’s public shares cashed out, and Agility now expects about $465 million gross, including $100 million already raised. Judge it on the first quarterly report, not the debut.
Washington calls imported robots a security risk
No tariffs yet. A presidential proclamation on 28 October found that imports of robots, industrial machinery and their parts threaten US national security, and gave trade officials 180 days to negotiate and report back. Pricing an automation project for 2027? Ask suppliers, in writing, who pays if a duty lands.
Siri opens up to every app
Apple released a developer preview yesterday that lets any app hand its actions to Siri AI, so people can check a delivery or book a table in their own words. App Store submissions open in the spring, and Apple expects customers to get the features in a few months. It’s currently unavailable on iPhone in the EU. Run an app? List the five things customers ask it most.
Anthropic, twice, on checking the checks
FlashProve, out on Monday with researchers from EPFL and MIT, attaches a cryptographic proof that a reply came from the model a provider claims to run. Prototype only: on Llama 4 Maverick, proving took 70 to 95 times the GPU time of serving it with vLLM. And in October research, datasets written by Claude agents cleared a row-by-row monitor’s acceptance threshold yet still made a never-deployed test model misaligned.
TOOL OF THE WEEK
The Wayback Machine
Step 2 of the prompt, made painless. Free, not new, not AI. The browser extension (Chrome, Firefox, Safari, iOS, Android) tells you when a page you land on is missing and a saved copy exists, and Save Page Now keeps a permanent copy of any page you cite, so your own source can’t vanish later.
The catches: one page at a time, some sites block it, and a report behind a form stays there. An archived copy proves a page said something once, not that it was right.
Verdict: install the extension, and archive every page you cite in anything a client will read.
TIP OF THE WEEK
Write the sample into the sentence
Most bad numbers aren’t wrong. They’re about somebody else. Among the changes Trakkr logged on statistics that did reach a study, only about a quarter were to the digits. The rest were context: what was measured, the year, the source, who was asked, the place.
So stop writing ‘62% of buyers prefer…’. Write ‘In a 2025 survey of 300 UK finance directors by [firm], 62%…’. Anyone who copies you copies the limits too.
The limit: sentences get longer. On a slide, keep the headline short and put the sample in a footnote on the same slide.
Pro move: ask your assistant to rewrite every number in a draft with sample, date and source inline, and to mark any it can’t fill as [UNKNOWN]. Those blanks are your trace list.
YOUR MOVE
Trace one number before you send it
Pick one statistic an AI gave you that’s in something going out this week. Run the prompt on it, then act on the verdict.
SOURCED: cite the study, not the page that quoted it. CHANGED: use the original’s figure and context. CAN’T CHECK: if it’s a side point and you can see the publisher’s own statement, write ‘[publisher] reports…’. Otherwise, and always when it carries a headline or a recommendation, cut it. NOWHERE: cut it.
REPLY WITH ONE LINE
Which verdict did your number get, and why? For example: ‘Pitch deck, market-growth figure: CHANGED, the study was from 2019 and US-only.’ Leave out client names. I read every reply.
Know someone whose slides say ‘94% of first impressions’? Forward them this.
Disclosure: the original draft used Claude. Anthropic, which makes Claude, appears above; Claude sourced 38% in Trakkr’s test.
P.S. Trakkr’s own AI judge, working alone, rated 23% of statistics sourced. Hand checks put it at 40%. Checkers miss in both directions, which is why the companion checks the checker:
iPROMPT / 160 /
A deal has two sides. Your stack only sees one.
Aligned shows you the other half. See what buyers actually do between meeting. See what they open, revisit, and share; so you know the exact next step to move things forward. And your champion gets everything they need to win internally when you're not in the room.