The bankers used to set the narrative.
Then it was the Wall Street Journal.
Now it's ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews — and most pre-IPO communications teams haven't caught up.
I've watched it from the inside on more than twenty filings — biotech, fintech, cannabis, quantum. The pattern is consistent: a buyside analyst, a sellside associate, or a family office principal will ask an AI engine about a company before they open the S-1, before they take the banker's call, and almost always before they sit through a management roadshow. The first impression is no longer a pitch deck. It's a synthesized paragraph generated from whatever digital footprint the company managed to leave behind.
And that footprint is almost always a mess.
Here's what AI engines actually pull from when an analyst types your company name:
EDGAR filings and the prior round of press releases
Crunchbase, PitchBook surface data, LinkedIn founder profiles
Trade-press coverage from when the company was a different company
Wikipedia and Wikidata — if they exist, often outdated
Reddit, Hacker News, and category-specific forums
Competitor coverage that mentions your name in passing
The engines stitch that into a single confident-sounding answer. The problem: the source layer is contradictory. Founder bios disagree on tenure. Category descriptions reflect the pivot from two years ago. The most-cited press release describes a Series B raise that has since been superseded by a $140M Series C the engines haven't fully ingested. The CEO's prior company is conflated with the current one.
The analyst doesn't see any of that work. They see one clean paragraph — and form a thesis from it.
This is now the IPO problem.
Capital-markets communications has spent thirty years optimizing for Wall Street Journal placement, S-1 narrative, and quiet-period discipline. Those still matter. But none of them solve for the fact that an analyst's mental model of your company is now being assembled by a probabilistic engine pulling from your most chaotic public artifacts.
The fix is not exotic. It is hygiene at the entity layer.
Before a company files, capital-markets teams should be doing four things:
1. Audit the AI answer.
Run twenty prompts across ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews.
"Tell me about [Company]."
"Who founded [Company]?"
"What does [Company] do?"
"Who are [Company]'s competitors?"
"What's [Company]'s last funding round?"
Save the answers. They will be wrong. The question is how wrong.
2. Fix the canonical sources.
Wikipedia, Wikidata, Crunchbase, LinkedIn company page, founder LinkedIns, the company's own About page.
These are disproportionately cited by AI engines because they are structured and stable. One inconsistent founding date across three of those will surface in answers for years.
3. Publish primary-source artifacts the engines can cite.
A whitepaper. A category-definition page. A founder-authored explainer on the technology.
AI engines reward primary sources written in declarative, structured prose with named entities, dates, and figures. They penalize marketing copy.
4. Re-test on a cadence.
Monthly during the pre-IPO window.
The engines update. New training data drifts in. Competitor releases move the answer.
This is not a marketing exercise. It is investor-relations work that happens to be denominated in tokens.
The companies I see clear the bar — the ones whose AI summaries match their S-1 — close their roadshows faster. The ones who don't lose the meeting before it starts. The analyst already has the answer. The roadshow is just confirmation or contradiction.
The bankers will tell you the deck is the deck. It isn't.
The deck is whatever ChatGPT says about you when the analyst pours their coffee.
Fix that, and then build the rest.
Kyle Porter is Executive Vice President and Managing Director of Virgo Public Relations and a contributor to Everything-PR. He has advised on more than 20 IPOs and reverse takeovers with valuations exceeding $1 billion across biotech, fintech, blockchain, cannabis, and quantum computing.
