To AI search, a gated whitepaper is mostly invisible. The crawlers that feed ChatGPT's and Claude's answers don't fill in forms, so the only part of your report an engine can quote is the landing-page teaser. Ungating everything isn't the fix. Publish the citable layer (findings, numbers, method) as HTML and keep the full asset gated. In one small July 2026 test, adding an HTML evidence page raised a PDF's citation rate from 35.4% to 60.4%.
A year ago, losing that visibility would have mattered less. In G2's April 2026 survey of 1,076 B2B software buyers, 51% said they now start software research with an AI chatbot more often than with Google, up from 29% in G2's 2025 report. The 6sense 2025 buyer experience study found buying groups have already filled four of five shortlist spots on Day 1 and first contact a seller 61% of the way through the process. Taken together, an engine is often helping draft the shortlist before anyone reaches your form. Your best research, the asset built to earn that shortlist spot, sits behind a form that no automated system will fill in.
What an AI engine actually sees from a gated report
It helps to separate the bots, because they do different jobs. OpenAI and Anthropic each run three kinds of crawler: one for model training (GPTBot, ClaudeBot), one for search (OAI-SearchBot, Claude-SearchBot), and a fetcher that fires when a user asks a question (ChatGPT-User, Claude-User). The second and third decide whether your page can appear in an answer. OpenAI's bot documentation is blunt about it: "OAI-SearchBot is for search," and sites that opt it out won't appear in ChatGPT search answers.
None of these bots treats your form as an invitation. Anthropic's crawler documentation says its bots "respect anti-circumvention technologies" and states: "we will not attempt to bypass CAPTCHAs for the sites we crawl." OpenAI's page says nothing about logins, forms or paywalls either way. That's the default: a crawler fetches what a public URL serves and moves on.
So what does a gated report contribute? Whatever is in the landing page's HTML. Usually that's a headline, three "In this report you'll learn…" bullets and the form. "Discover the trends reshaping B2B buying" is a promise, not a fact, and an engine can't cite a promise as evidence. As far as the answer is concerned, the survey data behind the form doesn't exist.
Three details that make the teaser even thinner
The first is rendering. Vercel and MERJ studied AI crawler traffic and reported in December 2024 that "none of the major AI crawlers currently render JavaScript", naming the crawlers of OpenAI, Anthropic, Meta, ByteDance and Perplexity. ChatGPT's and Claude's crawlers fetch JavaScript files but don't execute them. Gemini is the exception because it uses Googlebot's rendering infrastructure. Bot behaviour may have changed since that study, but if your landing-page builder injects the teaser client-side, those crawlers may see less than the teaser: an empty shell with a form embed.
The second is markup. Teams sometimes hope schema will fill the gap by describing the report in JSON-LD even when the prose is thin. Ahrefs tracked 1,885 pages that added JSON-LD against matched controls. ChatGPT citations moved +2.2% and AI Mode +2.4%, both "statistically indistinguishable from zero." Ahrefs also cites a searchVIU test in which "every system extracted only visible HTML content." The findings need to be on the page as visible text and tables. Describing them in markup doesn't count.
A third detail is specific to demand-gen stacks. Anthropic notes that robots.txt has to be set per subdomain. If your landing pages live on a marketing automation subdomain such as go.yourcompany.com, that subdomain has its own robots.txt and possibly its own CDN rules, both easy to overlook. Our guide to auditing robots.txt, CDN rules and AI crawler access walks through it, and the free AI Crawler Access Checker live-fetches a URL as the retrieval bots to catch blocks that robots.txt doesn't show.
All of this assumes the gate holds. Many don't.
When the gate leaks: wall or sieve
A gate fails in one of two ways. A wall keeps everything out, crawlers included, and the report is invisible. A sieve looks like a gate to humans but lets the content through to machines. A sieve sounds like a lucky accident. It usually isn't.
One common way a gate turns into a sieve is the file URL. In many setups, the form guards the link to the PDF, not the PDF itself: the thank-you page or follow-up email points to a file at a public address. Once that address shows up on any crawlable page (a partner's resource list, a community thread, a forwarded email someone posted), it can be fetched like any other public file. The other way is the JavaScript "soft gate": the full text ships in the HTML and a modal hides it until the form is submitted. Crawlers that don't execute JavaScript never see the modal, only the text.
What happens when a leaked PDF gets cited? OtterlyAI's dead-citation study analysed 20,092,769 URLs cited by seven AI engines over one month. PDFs were about 1.69% of them (338,691 files), and 10.66% of those were already dead. Also: "77% of cited PDFs carry no date in the file name at all." Picture your 2026 benchmark report cited as final_v3.pdf with no date, no context, no link back to the landing page, and no form between the buyer and the content.
A wall loses the citation. A sieve loses the lead and the framing, and sometimes the date. Neither one is a strategy.
"Why not just let the bots through?"
Someone will suggest serving crawlers the full report by user agent and keeping the form for humans. Be careful. Google's paywalled content documentation says Googlebot needs access if you want paywalled sections indexed. It also says the paywall structured data "helps Google differentiate paywalled content from the practice of cloaking," and cloaking "violates spam policies." Without that markup, serving bots something different from what users get is the textbook cloaking pattern. Non-Google engines haven't published an equivalent mechanism at all. There's a safer route.
The evidence for a public citable layer
The better option is a two-layer model: a public HTML page carrying the report's findings, with the full asset still behind the form. One test speaks to it directly, and it's worth knowing in detail, including its limits.
In July 2026, MaxAEO's Chris Han ran a 384-answer test of whether AI engines cite PDFs and videos. It covered eight surfaces: ChatGPT Search, Gemini, Perplexity, Claude with web search, Copilot, Grok, Google AI Mode and AI Overviews. Each asset was tested two ways: published natively on its own, and published with "a crawlable HTML page containing its publisher, summary, method, material limitation," plus a reviewed transcript or textual finding linking to the original.
The results:
- The PDF alone earned supporting citations in 17 of 48 captures (35.4%). With an HTML evidence page beside it, 29 of 48 (60.4%).
- Across all four formats, native-only assets were cited in 34 of 192 captures (17.7%). With evidence pages, 109 of 192 (56.8%).
- Where the citations landed: 82 of those 109 (75.2%) went to the HTML page, not the native asset.
The last number matters most for demand gen. Assets weren't just cited more often once they had an HTML companion. Across all four formats, three citations in four went to the companion itself. The HTML page is what gets cited, so the HTML page should carry your brand, your numbers and a path to the form.
Han is candid about the limits: one English-language domain, four new assets, a US locale and three days of captures (July 8–10). Two runs per prompt are "insufficient for statistical independence," the results are "directional first-party observations," and the evidence page added both a new URL and new text, so the two effects can't be separated. MaxAEO is also a vendor in this space. Treat the test as the best direct signal available, not as a law.
A second line of evidence explains why findings and numbers in particular get lifted. The GEO paper from Princeton and IIT Delhi researchers (KDD 2024) found that optimization methods "can boost visibility by up to 40%" in generative engine responses. By our reading of the paper's Table 1, adding statistics raised the position-adjusted word-count metric by roughly a third over the unoptimized baseline, adding quotations by more, while keyword stuffing slightly lowered it. The paper also found that "lower-ranked websites, which typically struggle for visibility, benefit significantly more from GEO." With statistics added, a fifth-ranked source's visibility rose 97.9%. That was a lab benchmark on 2023-era engines, not live ChatGPT. Still, its direction matches the MaxAEO test: engines reward passages that state a specific, attributable number.
The practical point is that a findings page isn't a marketing summary of the report. It's the report's evidence, written in sentences an engine can lift on their own. Engines retrieve and score passages, not whole documents, so each finding needs to work without the paragraph around it. Our explainer on how AI engines chunk pages for passage retrieval covers what that means for section structure.
You already know this model: Gartner, Forrester, HubSpot
If "publish the findings" sounds like giving away the store, look at how the biggest gated-research businesses handle their own reports.
Gartner sells its research. Its May 20, 2026 press release still publishes a complete citable layer on one public HTML page: the headline finding ("Sixty-nine percent of B2B buyers prefer to validate AI-generated insights with sales reps"), the sample (645 B2B buyers), the field dates (August through September 2025), secondary findings (45% used GenAI; buyers used an average of seven information sources on a recent purchase), and a named analyst. That page travels. Skyword's July 2026 article on ungating content for AI search cites the 69% figure as a 2026 Gartner survey. The number travels without the report.
Forrester follows the same pattern. Its paid State of Business Buying, 2026 report comes with a free press release that states, for example, that "the typical buying decision now includes 13 internal stakeholders and nine external influencers" and that procurement professionals are decision-makers in 53% of business buying cycles. Analyst posts such as this one on zero-click buying carry more survey figures, including that buyer AI use "grew five percentage points, to 94%."
HubSpot is the closest match to a demand-gen program. The State of Marketing report sits behind an Access Now form with required fields for first name, last name, email, website URL and number of employees, plus reCAPTCHA. Alongside it runs an ungated HTML post that sets out findings from 1,500+ global marketers in plain text. One example: "68.2% of marketers say they understand how to use AI in marketing, compared with 47% in 2025." The post repeats a "Download Free State of Marketing Report" call to action that leads back to the form. The findings are public, the full asset is gated, and the public layer feeds the gated one.
This is the framing to take into the budget meeting. You aren't being asked to give away the report. You're being asked to do what Gartner does. The organizations that monetize research most aggressively publish the number, the sample and the method, and still sell the report.
What to ungate and what to keep gated
That gives you a working rule. If an AI engine could answer a buyer's question from your report, you want it to answer with your number, under your name. If the buyer needs to do something with your data (segment it, apply it, benchmark against it), that's what the form is for.
Skyword's Casey Nobile states it more briefly: "make your expertise public, and make its application exclusive." Google's own guidance points the same way. Its flexible sampling documentation describes a "lead-in" model that shows a portion of content without the full piece. On showing the first few sentences above a paywall, it says: "We think this is a good practice."
| Part of the report | Public (HTML) | Gated | Why |
|---|---|---|---|
| Headline findings, one per sentence, with numbers | ✓ | These are the passages engines quote | |
| Sample size, field dates, geography, method, limitations | ✓ | Makes the number attributable; mirrors MaxAEO's evidence page | |
| Key charts, redrawn as HTML tables | ✓ | Engines extract visible text, not images | |
| Definitions and frameworks the report introduces | ✓ | Category language you want associated with your brand | |
| Full narrative and analysis | ✓ | The depth a serious buyer will trade an email for | |
| Segment cuts (by industry, company size, region) | ✓ | The "how do I compare?" layer: high intent, high value | |
| Raw dataset, templates, calculators, benchmarking tools | ✓ | Application, not expertise | |
| Analyst briefing or webinar | ✓ | Live access is a natural gate | |
| Anything containing customer data | ✓ | Not a visibility decision at all | |
| Assets syndicated under a pay-per-lead contract | ✓ | MQL volume is the contracted deliverable |
The keep-gated column draws on practitioners. Skyword's ungating framework keeps gates on interactive tools, live experiences, bespoke deliverables and transactional steps, and offers one admission test: does gating this asset cost a citation you would otherwise earn? Deepak Gupta, whose full-ungating experiment appears below, still keeps proprietary-data assets, sales-qualifying assets such as an RFP template, and partner campaigns paid per MQL behind a form.
The citable layer, specified
A findings page fails when it reads like a teaser. Build it to this spec:
- An answer-first opening. Two to four sentences that state the single most important finding with its number and scope.
- One finding per sentence, number included. A sentence that gives the exact share of a named buyer group beats "most buyers…" every time, and each sentence should stand on its own if it's lifted.
- A method box. Publisher, sample size, who was surveyed, field dates, geography, and one honest limitation. That's what turns a number into a citable claim, and it's the structure MaxAEO's evidence pages used.
- Tables, not chart images. Redraw the two or three key charts as HTML tables.
- A visible "last updated" date. OtterlyAI's finding that 77% of cited PDFs carry no date in the file name is a reminder that undated evidence travels badly.
- A clear path to the full asset. Name what's behind the form: the segment cuts, the benchmark tool, the full analysis. Don't use a vague "download for more insights."
- Server-rendered HTML. Given the Vercel findings, the findings themselves should be in the initial HTML response, not loaded by JavaScript.
Selling it to a CMO graded on MQLs
The hard part isn't the page. It's the conversation about the number your team is paid on, so start with the most uncomfortable data point.
Deepak Gupta went further than this guide recommends. In January 2025 he ungated all 16 of his e-books and tracked the results for twelve months. MQLs fell 71% year over year and SQLs fell 12%. Opportunities were flat, and closed-won rose 4%. AI citation share on his tracked queries across ChatGPT, Claude and Perplexity rose 4.1x, and branded search grew 38%. AI-referred buyers came in at 1.8x the deal size of form-captured MQLs, with a 22% shorter sales cycle. His summary: "Models cannot cite what they cannot retrieve."
Treat that as one founder's anecdote: no methodology, control group or attribution rules are published. It's still the most useful case for this conversation because it shows the likely shape of the trade, with the MQL count dropping sharply while revenue holds. A partial ungate (findings public, report gated) is a much smaller bet. Be straight about what nobody knows yet, too: no one has published a controlled before-and-after on form-fill rates after adding a findings page. Gartner, Forrester and HubSpot running both layers suggests the form survives, but it isn't proof.
Three more data points help reframe the KPI:
- AI-influenced buyers still come to sales. Gartner's survey found 69% of B2B buyers prefer to validate AI-generated insights with sales reps. Getting cited doesn't route around your pipeline. It changes who arrives at it and how far along they are.
- Inclusion shapes preference. In G2's survey, 85% of buyers think more highly of a vendor when AI includes it in an answer, and 69% chose a different vendor than they initially planned because of a chatbot's recommendation.
- The clicks are few but strong. On Ahrefs' own site, AI search visitors drove 12.1% of signups from 0.5% of traffic, a conversion rate 23x that of traditional organic. The same post roughly estimates AI search users click links 75% less. That's one company's data, not a benchmark, but it fits the idea that the value is influence and shortlisting more than sessions.
The pitch to a CMO isn't "trust me, MQLs don't matter." It's a bounded pilot with pre-agreed metrics, on a handful of assets, where the downside is capped and the result is measured.
Measure the trade before you move
Skyword suggests ungating three to five information assets and watching them for 60 to 90 days, tracking AI citation frequency, AI referral quality, sales-cycle length and close rate. For the partial model, the pilot looks like this:
- Pick the prompts the report should win. For a benchmark report, these are questions like "what's a typical [metric] for [category]" or "how many stakeholders are in a B2B buying decision". Use the research-stage questions your report actually answers, not your brand name.
- Baseline before you publish. Record mention rate (does the engine name you?) and citation rate (does it link your domain?) per engine for those prompts. For a gated report, expect citation rate near zero. That's the wall at work.
- Ship the findings page to the spec above, server-rendered, on a crawlable path, linked from the report's landing page and from your blog.
- Confirm the bots can fetch it. Server logs will show whether OAI-SearchBot or Claude-SearchBot requested the page, but as our piece on reading AI crawler logs without confusing crawl volume with citations explains, a fetch shows access, not use.
- Track the gated form separately. Your CMO will ask about form fills on the full asset. Report them alongside opportunities and closed-won, so a dip in raw fills is read in context.
- Mark the publish date and wait for the window. AI answers are nondeterministic, so one check after launch tells you little. You need a trend across repeated runs.
This is where Geoptimizer fits. It runs your tracked buyer-intent prompts on ChatGPT, Gemini, Claude and Grok with web search on, and reports mention rate and citation rate per engine on a 7-day rolling window with a confidence band, which answers the "was that lift just noise?" question. The day the findings page goes live gets marked on the trend chart, the same way a robots.txt fix does, so any later change has a date to compare against. Watch citation rate: a findings page that works shows up as engines linking your domain, not just naming your brand.
FAQ
Can ChatGPT or Claude read content behind a lead form?
Not through their search crawlers. OAI-SearchBot and Claude-SearchBot fetch what a public URL serves, and Anthropic states it won't attempt to bypass CAPTCHAs. Agents are a separate case: OpenAI's ChatGPT Atlas browser, launched October 21, 2025, has an agent mode whose launch demos included filling out forms. An agent acting for one user may submit your form, but that doesn't put your report into anyone else's answer.
Will publishing the key findings reduce form fills?
Possibly, and no one has published a controlled test on it. What's observable is that Gartner, Forrester and HubSpot publish headline numbers, samples and dates while still selling or gating the full report. Keep the high-intent layers (segment cuts, tools, raw data) behind the form, and measure fills during a pilot rather than assuming either way.
Is a JavaScript soft gate a good compromise?
Usually not. If a modal hides full text that's already in the HTML, non-rendering crawlers read everything: visibility without lead capture, plus a cloaking risk in Google without paywall markup. A separate findings page is cleaner.
How long should a gated-content pilot run?
Skyword recommends 60 to 90 days on three to five assets. That's long enough to see citation trends settle across repeated runs and short enough to report back within a quarter.
The short version
A form is a wall to the bots that assemble AI answers, and a leaky form is a sieve that hands engines an undated PDF with no lead capture. The model that works is the one the big research firms already use: publish the findings, numbers and method as citable HTML, and keep the analysis, segment cuts and tools gated. The evidence is directional, so baseline the prompts your report should win, ship the findings page, and watch citation rate per engine. Geoptimizer's plans start free, with all four engines included, so you can begin tracking those prompts before the findings page goes live.