A plant manager at a 300-employee contract manufacturer in the Midwest needs a new vendor for industrial fasteners. Her current distributor missed two deliveries last quarter and the line went down twice. In 2022, she would have called a few peers, asked her purchasing lead, maybe walked the floor at a regional trade show. In 2026, she opens ChatGPT. She types something like, "We are a mid-sized manufacturer in Ohio. We need a fastener distributor that can hold safety stock and ship same-day on emergency orders. Who should we look at?" She gets back four vendors with a paragraph on each. She clicks into two of them. That is her shortlist.
The entire discovery stage you used to call the top of the funnel just happened, and you are not in it.
Forrester's 2026 State of Business Buying, based on nearly 18,000 buyers globally, found that 94% of B2B buyers now use AI somewhere in their purchase research. A separate study from 2X, an AI visibility index that benchmarked 70 B2B companies, found 96% of them were invisible in early-stage AI-driven buyer discovery. They only appeared once a buyer typed a branded query. By then, the shortlist was already set.
So this is not a missed deal. It is a deal that never existed for you. There is no closed-lost report. No anonymous web traffic. No "we evaluated and chose someone else." Just absence.
The Same Operational Extraction Problem Exists In PO Processing
A PO arrives through email. Someone reads it, validates the information across systems, and rekeys the data into the ERP.
That is not order management. That is manual data extraction.
Y Meadows Starter Web App automates the workflow - extracting order data directly from customer orders and delivering an ERP-ready output file in minutes.
Less manual work. Fewer errors. Faster order processing.
Why Marketing Can't Fix This Alone
The natural reaction is to send this problem to marketing. Partly fair. Marketing owns publish. They will hire an AEO consultant, add schema markup, and start tracking citations in Perplexity. That work matters. But it will not move the needle if the underlying material is wrong.
AI search engines are not impressed by adjectives. Phrases like "trusted by leading manufacturers" do not get cited. What gets cited is specificity. A real industry, a real problem stated in the buyer's own language, a real outcome with a number. "Reduced unplanned line-down events from 6 per quarter to 1 per quarter at a 200-employee contract manufacturer in Ohio, by holding safety stock on 14 fastener SKUs" gets surfaced. "Our customers love our streamlined workflows" does not.
Most B2B marketing pages read generic because they were written from positioning docs, not customer reality. The specific stuff lives in account manager check-in calls, in customer service email threads, in notes after a site visit, and in the customer email someone forwards around the office with the subject line "thought you'd want to see this." Marketing has none of that. You have all of it.
You Are Already Sitting On The Source Material
Account managers and customer service reps are the only people in your company who routinely hear customers describe how they actually use what they bought from you. What goes right. What almost went wrong and got saved. The new use case nobody at headquarters knows about. A 30-minute check-in call with a long-time customer has more usable material than a year of marketing surveys. And the email thread where a customer explains why your substitute SKU saved their shift gives you the actual story, not the abstracted version.
The problem is that this material never leaves the system it was captured in. It sits in a Teams recording nobody re-watches. It sits in a customer service inbox that gets archived monthly. It sits in a folder of visit notes only the account manager ever reads. Nobody is pulling it out and turning it into anything. So marketing writes from the positioning doc, again, and the cycle continues.
One AI workflow this week: a Claude Project that turns raw customer-facing conversations into structured customer stories that are usable downstream. The trick is not the project. It is the output schema.
Build It: The Customer Story Extraction Project
Setup is short. Create a Claude Project. Call it "Customer Story Extraction." Add the following as project knowledge:
- Last quarter's call transcripts from Teams or Zoom: account manager check-ins, customer site visits, problem-solving calls. Top 30 by length is enough for a first pass.
- Customer service email threads from the past 90 days, especially the ones that ran long because the customer described their situation in detail.
- Notes from customer site visits or account review meetings, even the rough ones.
- Anywhere else customers describe their setup, their constraints, or what they got out of working with you, in their own words. Skip anything your team wrote about the customer.
The custom instructions for the project should tell it to output one story per source document, using a fixed schema. The schema is the part that matters more than anything else in this issue, so read it twice.
For each source document, extract one customer
story using exactly these fields. If a field
cannot be filled from the source, write
"NOT IN SOURCE." Do not invent details.
1. CUSTOMER PROFILE: industry, company size
(employee count or revenue band), region,
role of the buyer.
2. TRIGGER: the specific event that made them
act (volume spike, new facility coming
online, supply disruption from previous
vendor, equipment change, regulatory
requirement, new ops or purchasing lead).
3. PROBLEM IN CUSTOMER'S OWN WORDS: a direct
quote of 1-2 sentences describing the
problem before they engaged with us.
4. WHAT THEY TRIED FIRST: alternatives
considered or attempted before our solution.
Name competitors only if the customer named
them in the source.
5. DECISION CRITERIA: the 2-3 things the
customer explicitly said they cared about.
6. WHAT CHANGED: the specific outcome, with
a metric if one is in the source. No metric
means write "no metric stated."
7. CUSTOMER QUOTE: an attributable sentence
in the customer's voice describing the
outcome or impact. Verbatim from the source.
8. TIME TO VALUE: how long until the customer
saw the result.
9. CAVEATS: where this story may not apply.
Industry-specific aspects, scale-specific
limits, edge cases the customer mentioned.
Output one story per source. Do not paraphrase
quotes. Do not smooth the customer's voice into
marketing language.
Each field is doing a job in how AI surfaces the eventual story.
- Profile fields let AI match the prospect's context. A buyer asking "what works for a 200-employee distributor" gets your story if the profile actually says 200-employee distributor. Vague identifiers like "mid-market" do not match anything.
- Problem in own words gives AI the actual language buyers use to search. Buyers do not type "industrial distribution partner optimization." They type "we keep running out of bearings on second shift and the line goes down." Capture that.
- What they tried first pre-handles competitive comparison prompts. When a buyer asks AI "is X better than Y," the answer references content that mentions both.
- Decision criteria match the most common research prompt pattern. "What should I look for in an order management vendor" returns lists. Your stories should be the source of those lists.
- Outcomes with metrics are what answer engines cite. Numbers get pulled into responses. Adjectives do not.
- The verbatim quote is the attribution unit. AI systems prefer cited claims over uncited ones. A direct quote with a real name behind it carries more weight than the same point in your voice.
- Caveats are the counterintuitive one. Marketers want to scrub these out. Don't. AI gives more weight to sources that acknowledge limits, the same way you trust a Yelp review more when it lists a negative alongside the positives.
An Honest Take
The companies that win at AI search in 2026 will not be the ones with the best SEO consultants. They will be the ones whose sales and CS teams have a working system for extracting what their customers actually said. Most companies will not bother. They will commission another round of marketing case studies, written from positioning docs, that read like every other vendor's case studies. And they will wonder why AI keeps recommending their three biggest competitors. The asymmetry is real. The material already exists in your transcripts and tickets. Whoever extracts it first builds the citation depth before anyone notices. By the time the laggards catch on, you have a 6-month lead in the only training data and live retrieval surface that matters. That is the moat. Not better positioning. Better source material, systematically pulled.
What to Watch Out For
- Hallucinated metrics. AI projects, even with grounded instructions, sometimes invent numbers that round nicely. Always check the metric against the transcript. If the source says "way faster" and the output says "40% faster," that 40% came from nowhere. Cut it.
- Smoothed-out quotes. The model wants to make the customer sound articulate. The customer was on a Tuesday afternoon call with a half-eaten sandwich. The rough version is the credible one. Compare the output quote to the transcript and revert anything that got polished.
- Customer permission. Not optional. Nothing with a customer name, quote, or identifying detail leaves your company without sign-off. Set up the permission flow before you start producing stories, not after.
- Sample size. One quarter is the minimum. One month gives you anecdotes, not patterns. If you only have a handful of detailed sources, run the project anyway, but treat the output as a pilot, not a corpus.
- Story decay. Anything older than 18 months should be revisited or retired. Buying behavior in 2025 is not buying behavior in 2026. The product is also different. So is the competitive set.
The Handoff to Marketing
Run the project. Review the output. Hand the structured stories to marketing. That is your job done.
Marketing's job is to publish them under a real customer name (with permission), wrap each one in the right schema for AI crawlers to parse, and place them on pages that get indexed. That part is real work and not trivial. But marketing cannot do any of it without the structured source material you just produced. Without you producing, they publish positioning. Positioning does not get cited.
Set a cadence. Once a quarter is the floor. Once a month is the goal. The first batch is the hardest. After that, the project gets faster because you have a working schema and the team knows what good output looks like.
The Bottom Line
94% of your buyers research with AI. 96% of B2B companies do not show up in the first answer. That is not a marketing problem you can outsource. It is an extraction problem. The raw material lives in your account manager check-ins, your customer service emails, and your visit notes. The pipeline from there to anywhere usable is what is missing.
Build the project. Use the schema. Run it on last quarter's calls and emails. Review the output for hallucinated metrics and smoothed quotes. Get customer permission. Hand the result to marketing. The first stories go live in a quarter. Citation lift takes another quarter after that. The companies starting now are building a 6 to 12 month lead that gets harder to close every month.
Your Team Should Not Be Retyping Customer Orders In 2026
While your competitors are automating intake end-to-end, your best people are still copy-pasting order data into your ERP.
Y Meadows automates manual order entry - from incoming customer orders through validation and ERP entry.
Takes 20 minutes to show how it works.