Can B2B buyers really detect AI-generated copy?

Yes, and the evidence is no longer anecdotal. TrustRadius (2025, n=2,058 verified buyers) reports that 62% of frequent AI users always or very often fact-check the marketing content they read. Forrester's 2026 B2B Predictions show 19% of buyers exposed to genAI in their decision process felt less confident about their final purchase because of unreliable AI output. The pattern-recognition curve in cold outbound is now under three seconds. Buyers do not need to articulate why a piece of copy reads as AI. They down-weight the source and move on.

Is AI-as-editor the same risk as AI-as-author?

No, and the distinction is the most important operational rule in this guide. Kirk and Givi (JBR Vol. 186, 2025) explicitly carve out that the trust loss is attenuated when AI only edits a communication rather than authors it. The defensible workflow is to write the first draft as a human, use AI to tighten, condense and pressure-test, and then ship under a named byline. The damaging workflow is to ask the model for a draft, lightly skim, and publish.

Does this apply to factual product copy and documentation?

Largely no. The Kirk and Givi finding is also explicitly attenuated for factual messages, including pricing pages, technical specifications and comparison tables. Spec sheets, changelogs and API documentation can be AI-assisted heavily without measurable trust loss. The exposed surface is emotional and brand content: founder posts, About pages, customer stories, cold outbound, newsletters and LinkedIn.

What is the single highest-leverage edit a marketing team can make this week?

Build a banned-word file inside the brand voice guide, add the Originality.ai short list (delve, tapestry, multifaceted, leverage, robust, seamless, harness, unlock, in today's fast-paced world, it's worth noting), and forbid them in any AI-assisted draft without explicit override. Co-occurrence is the signal. Three banned words inside a 500-word passage raises AI-detection probability by 35 to 50 percentage points across modern classifiers. The cost of this change is one document and one editorial rule.

How do I audit my team's existing content without rewriting it all?

Run the 10-point AI tell audit against a representative sample of the last quarter's published content. Score each piece on banned-word density, structural ticks (it's not just X, it's Y), em-dash density, burstiness, opener and closer formula, claim specificity, named human presence, real numbers, stated point of view and byline accountability. Anything scoring under 6 out of 10 goes into a rewrite queue. Anything under 3 should be unpublished or noindexed.

Which channels demand the most human attention?

In order: cold outbound, founder and CEO LinkedIn, customer stories, About pages and newsletters. These are the surfaces where AI-as-author causes the most measurable trust loss. The channels where AI-as-editor is safest are technical documentation, changelogs, pricing pages, transactional emails and paid-ad creative variants.

Does Google actually penalize AI content?

Google penalizes AI slop, not AI assistance. The 2023 Search Central guidance, restated through the March 2024 and March 2026 core updates, rewards original, high-quality content that demonstrates E-E-A-T (expertise, experience, authoritativeness, trustworthiness) and de-ranks thin, scaled, undifferentiated content regardless of whether it was written by a human or a model. Sites with named authors, original research and demonstrated experience gained ground in both updates. Sites publishing programmatic AI SEO content lost majorities of indexed pages.

What KPIs prove the trust gap is closing?

Five leading indicators. Branded search volume on the company name and the founder name (rising). Reply rates on cold outbound (rising toward and through the 5.5% top-quartile threshold). Comment-to-impression ratio on LinkedIn personal posts (rising). Inbound demo requests citing a specific article or post (rising). And, on the lagging side, win rate against incumbents in deals where the buyer cited reading something the marketing team published (rising).

The Word 'Delve' Is Costing You Deals: Why B2B Buyers Already Know Your Marketing Is AI Slop

By Juan Beltrán · 2026-05-07

B2B buyers can already read the linguistic fingerprint of AI-generated copy, and they discount the brands shipping it. Kirk and Givi proved the trust loss in seven preregistered experiments. Originality.ai catalogued the tells across 3.3 million documents. 73% of B2B buyers trust peer recommendations over AI chatbots. The teams scaling AI without humanizing it are accumulating a quiet, unattributable trust debt that surfaces as a slow erosion in reply rates, branded search and shortlist inclusion. The fix is operational, not philosophical: AI as junior writer, a public brand voice guide, and a 10-point pre-publish audit on every draft.

Three identical posts

A few weeks ago I read three LinkedIn posts in a row that all began with the words "In today's fast-paced world." One was a CEO. One was a brand-new founder. One was a competitor. They had the same opening line, the same three-bullet middle, the same emoji-flecked closer. None of them had written a single word of it.

That moment crystallized what had been a quieter observation across the B2B marketing industry for almost two years. Generative AI did not start as a content problem. It started as a productivity miracle. We could finally turn a half-formed idea into a draft in thirty seconds. We could go from one ad variant to ten. We could ship.

Somewhere between the productivity miracle and the third quarter of 2025, the cost of that speed became visible. Buyers started fact-checking us more. Reviewers started asking us pointed questions about authorship. Detection scores on our blog content drifted upward. Reply rates on outbound went down. The voice we had spent years earning began to sound, on certain channels, exactly like everyone else's.

The thesis of this piece is direct. The trust gap created by AI-generated copy is now the conversion gap. Most CMOs will not see it until their next QBR, because their AI dashboards report activity (drafts shipped, variants tested, cost per asset down) and the damage is unattributable (slow erosion of reply rates, declining branded search, quiet absence from peer recommendations). By the time the lagging indicators turn, the trust debt is already compounded.

Part I. The fingerprint is real

Stylometric analysis is not new. The same statistical fields that attribute anonymous Renaissance manuscripts to Marlowe or Shakespeare have been retooled, in the past two years, against the output of GPT-4-class models. Three pattern families do most of the work.

The first is vocabulary. Originality.ai's analysis of 3.3 million human-written and AI-generated documents found certain words appearing at extreme frequency multiples in AI text. The word delve is the canonical example, appearing roughly 48 times more often in AI output than in human writing. Tapestry sits at 35 times. Multifaceted at 28. It's worth noting at 31. In today's digital age at 24. Single instances mean nothing. Delve is a real English word and competent humans use it daily. The signal is co-occurrence. Three or more such words inside a 500-word passage raises AI-detection probability by 35 to 50 percentage points across most modern classifiers.

The second is structure. GPT-4-class models, trained heavily on RLHF feedback that rewards balanced argumentation, default to a small set of patterns. Tricolons. Antithetical constructions ("it's not just X, it's Y" is now the single most reliable structural fingerprint of GPT-4-class output). Em-dash overuse. Sean Goedecke's token-frequency analysis found GPT-4o uses roughly 10 times more em-dashes than GPT-3.5. The arXiv paper The Last Fingerprint, March 2025, measured GPT-4.1 producing 9.1 em-dashes per 1,000 words even when explicitly instructed not to.

The third is rhythm, or what AI-detection literature calls burstiness. Burstiness measures the variance of sentence length across a passage. Human writers swing wildly. They write a fragment. Then a long, winding sentence that carries three subordinate clauses, an aside, and a punch at the end. AI writers cluster around a 15 to 22 word mean. ProofreaderPro.ai's 2025 analysis of 200 academic samples quantified the gap. Human-written academic text shows a sentence-length standard deviation of 8.2 words. GPT-4o averages 4.1. Claude is slightly better at 5.3. Plot the sentence-length sequence of any AI passage and you see a near-flat horizontal band. Plot a competent human writer and you see a jagged skyline.

The 12 words your team should ban this quarter, in no particular order: delve, leverage, navigate, landscape, tapestry, realm, robust, seamless, foster, harness, embark, unlock. Add to the list every time an editor flags a phrase as off-brand. The point is not to ban words. The point is to force a deliberate decision about each one.

Part II. The trust gap is now the conversion gap

Colleen P. Kirk (NYIT) and Julian Givi (West Virginia University), writing in the Journal of Business Research (Vol. 186, 2025, article 114983), report results from seven preregistered experiments. Their headline conclusion: when consumers believe emotional marketing communications are written by an AI rather than a human, positive word of mouth and customer loyalty are reduced. The mechanism is serial, mediated by perceived authenticity and a measurable response the authors term moral disgust.

Two carve-outs of immediate operational value. The effect is attenuated for factual messages such as pricing pages, technical specifications and comparison tables. The effect is attenuated when AI only edits the communication rather than authors it. Translation: AI as your editor is far safer than AI as your author. Your spec sheet survives disclosure. Your founder story may not.

The TrustRadius 2025 Bridging the Trust Gap report (n=2,058 verified buyers, 490 vendors) gives the operational picture. 72% of B2B buyers now report encountering Google's AI Overviews during research. 90% click through to the underlying source. The most frequent users of AI tools, the segment most marketers assumed would be most receptive to AI-generated content, are the most aggressive fact-checkers. 62% report always or very often verifying what they read.

Forrester's 2026 B2B Predictions add a more troubling figure. Among buyers exposed to genAI during their decision process in 2025, 19% reported feeling less confident in their final purchase decision because of unreliable AI output. That confidence drop converts directly into longer sales cycles, larger buying groups and more deals lost to incumbents. Edelman's 2025 Trust Barometer Special Report on Brands captures the upstream cultural shift. 73% of people say their trust in a brand would increase if it authentically reflected today's culture. Buyers are not asking marketers to be more polished. They are asking them to be more recognizable.

Part III. Your dashboard will not catch this

The activity-impact gap is the reason most marketing leaders will deny the trust debt until it is too expensive to fix.

AI dashboards report what AI is doing. Drafts produced. Variants tested. Cost per asset. Time saved. Every one of these moves in the right direction the moment you deploy a content team on a model. The damage moves in a different system entirely. Reply rates on outbound. Branded search volume. Inbound demo requests citing a specific piece of content. Comment-to-impression ratio on personal posts. Win rate against incumbents in contested deals. Net revenue retention in segments where the buyer reads marketing content before renewing.

Cost per asset down and reply rate down can co-exist for six quarters before anyone connects them. That is the half-life of the trust debt. The leading indicators to watch instead are the five named above, run as a monthly review against the prior six-month rolling baseline. Anything declining for two consecutive months in a stable spend regime is a signal.

The operator test I use is simple: would a named sales leader be proud to forward this asset to one strategic account? If the answer is no, the automation saving is probably being booked in marketing while the trust cost is being paid by sales.

Part IV. The 10-point AI tell audit

Run this against a representative sample of the last quarter's published content. Each item is one point.

One. Banned-word density. Three or more entries from the banned list inside a 500-word passage scores zero on this item.

Two. Structural ticks. Any "it's not just X, it's Y" construction, three or more tricolons in a single section, or repeated parallel-structure openings score zero.

Three. Em-dash density. More than 4 em-dashes per 1,000 words in a writer who has never historically used them scores zero.

Four. Burstiness. Sentence-length standard deviation under 5.0 words across the passage scores zero.

Five. Opener formula. Any opener beginning with "In today's", "In an era of", "As we navigate" or a hedged setup of an obvious truism scores zero.

Six. Closer formula. Any closer that lands on "ultimately", "in conclusion" or a balanced "while X, Y" scores zero.

Seven. Claim specificity. Any unsupported assertion not backed by a number, a named source or a dated event scores zero.

Eight. Named human. The presence of a real customer, employee or industry name in the body scores one. Generic "leading enterprises" scores zero.

Nine. Real number. The presence of at least one specific number not invented by the writer scores one.

Ten. Point of view. The piece argues for or against something specific, with at least one sentence the average competitor would not write, scores one.

Anything scoring 6 or above ships. Anything between 3 and 5 goes back to the writer with the failing items called out. Anything under 3 should be unpublished or noindexed.

Part V. The fix is operational, not philosophical

Treat AI as a junior writer, not a final author. The defensible workflow has five steps.

One. Brief the work as a human, in writing, with a named reader, a named customer reference, the desired action and the one belief the company holds that competitors do not. The brief is the prompt the model will never read.

Two. Generate the first pass with AI against the brief. Expect a 70% draft. Anything the model produces above 70% should raise suspicion that the brief was too generic.

Three. Humanize. This is the 40% tranche of the workflow that most teams underinvest in because it looks like editing. It is the highest-leverage human work in the entire content function. Strip the banned words. Break the parallel structures. Vary sentence length aggressively. Replace the abstract noun with a real customer name. Add the one number the model could not invent.

Four. Read it aloud. If you stumble on a sentence, the reader will too. Spoken-language and written-language work differently. AI defaults to written-language register on every channel. Newsletters, video scripts and cold emails all need to pass the read-aloud test.

Five. Ship under a named byline with a real photo and a real LinkedIn link. The byline is the trust contract. The Hustle's open-rate advantage over the industry average (35-45% versus 21%) is built on this single discipline. The byline is the unit of measurement.

The brand voice skeleton sits underneath the workflow. The first thing a public voice guide should declare is what the brand is not. Mailchimp's public guide opens with "we are not your weird internet aunt." That single negative-space line has done more to defeat AI-default voice inside Mailchimp than any prompt template. LLMs need exclusions more than aspirations. Document the variant per surface. Error messages warmer than legal copy. Transactional emails quieter than marketing emails. The voice flexes. The personality does not.

Part VI. Before and after

A typical AI-default cold email reads as follows.

Hi [First Name], I hope this message finds you well. I wanted to reach out because I noticed you're at [Company] and thought you might be interested in how our innovative platform helps forward-thinking teams unlock new efficiencies. We work with leading enterprises to streamline operations and drive measurable results. Would you be open to a quick discovery call to explore potential synergies? Best regards.

Twelve markers. "I hope this message finds you well." "Innovative platform." "Forward-thinking teams." "Unlock." "Leading enterprises." "Streamline." "Drive measurable results." "Discovery call." "Potential synergies." Industry average reply rate at 3.43% (Instantly, 2026).

The same outreach, rewritten for an elite-tier sender:

Maria. Saw your team posted three SRE roles in São Paulo last week. The pattern (LATAM hiring against a flat HQ headcount) usually means platform reliability moved up the board's agenda. We helped Toast cut MTTR by 41% in the same window. Worth 12 minutes Thursday at 14:00 your time?

One trigger (the SRE postings). One named customer (Toast). One real number (41%). One binary ask. No "leading platform." No "synergies." No greeting. Reply rates at this tier sit above 10% on platforms that can measure them. The AI-default sender is competing against this version, every day, in every inbox.

Part VII. Who is already winning

Mailchimp built a public, versioned brand voice guide in 2018, treats it as a living document, and revises it at least quarterly. Every contractor, every AI prompt, every junior copywriter operates against the same artifact. Customers describe the brand as having a personality, vanishingly rare in B2B SaaS.

Linear's changelog reads like an engineer who has shipped. No "we're excited to announce." Past tense, first person plural, specific. AI is used heavily for first drafts and bug-report triage. The output is then rewritten, not edited, by a human who has shipped software. The result reads like this person could have written it without AI. That is the test.

The Hustle ships a daily newsletter to 2 million plus subscribers at 35-45% open rates against an industry standard of 21%. AI in research and ideation. Every issue rewritten end-to-end by an editor on staff. Every byline a named human with a photo and a real LinkedIn. The byline is the accountability mechanism.

Anthropic publishes its own marketing. Engineering and research blog posts are written by named researchers. Customer stories run with real names and real numbers. The contrast with the AI competitors that ship the most generic-sounding content on the market is itself a positioning move. The company building the model writes its own copy. The signal is unmissable.

The uncomfortable conclusion

The brands that do not fix this in 2026 will not see a sudden drop. They will see a slow, unattributable erosion that their AI dashboards will explain away as market conditions. Reply rates will drift down a quarter at a time. Branded search will soften. Peer recommendations will quietly route to a competitor. The shortlist will be cut before the marketing team appears on it.

The authenticity premium compounds. The slop discount compounds too. The teams that learn to scale AI without losing the human voice that made them worth listening to in the first place will own the next ten years of B2B marketing. The teams that do not will spend those ten years explaining why their content metrics look great and their pipeline looks the way it does.

The full 54-page field guide, with the banned-word file, the brand voice skeleton, the 10-point audit, the channel-by-channel playbook, four full case studies and five before-and-after rewrites, is available below as a free download.

Research summary

Abstract

This field guide argues that the trust gap created by AI-generated B2B marketing copy is now the conversion gap. Three independent strands of evidence converge on the same conclusion. Kirk and Givi's seven preregistered experiments in the Journal of Business Research (Vol. 186, 2025) measure a serial moral-disgust response to AI-authored emotional marketing communications, with quantified declines in perceived authenticity, customer loyalty and positive word of mouth. Originality.ai's analysis of 3.3 million documents catalogues the detection-grade linguistic fingerprint, including the word delve at 48 times human frequency and a burstiness collapse from a human standard deviation of 8.2 words to GPT-4o's 4.1. TrustRadius, Forrester and Edelman quantify the buyer-side response: 73% peer-trust over chatbots, 62% always or very often fact-checking, 19% reduced purchase confidence. The fix is operational: AI as junior writer, brand voice as operating system, and a 10-point pre-publish audit applied to every draft.

Research question

Is the productivity uplift B2B marketing teams are extracting from generative AI being offset, on the buyer side, by a measurable, durable trust discount that suppresses pipeline before it ever appears in a marketing dashboard?

Methodology

Synthesis of peer-reviewed research (Journal of Business Research, International Journal of Advertising, arXiv preprints), industry benchmarks (Forrester, Gartner, Edelman, BCG, Content Marketing Institute, TrustRadius, ZenABM, Instantly, Originality.ai's 3.3 million document corpus), and first-hand operating experience across $500M+ B2B marketing portfolios. Where claims are contested they are reported as such. Where benchmarks differ across providers, ranges are given.

Key findings

Kirk and Givi (JBR Vol. 186, 2025): seven preregistered experiments demonstrate AI-authored emotional marketing reduces customer loyalty and positive word of mouth, mediated by perceived authenticity and a measurable moral-disgust response.
73% of B2B buyers (Reddit and SurveyMonkey, 2025, n=1,200) trust peer recommendations over AI chatbots when making purchase decisions.
Originality.ai's 3.3 million document corpus shows the word delve at roughly 48 times human frequency in AI-generated text, with tapestry, multifaceted and it's worth noting at 35x, 28x and 31x respectively.
Burstiness collapse: human academic prose shows a sentence-length standard deviation of 8.2 words. GPT-4o averages 4.1. Claude averages 5.3 (ProofreaderPro.ai, 2025).
TrustRadius (2025, n=2,058 buyers, 490 vendors): 72% of B2B buyers encounter Google's AI Overviews during research, 90% click through to source, 62% of frequent AI users always or very often fact-check what they read.
Forrester 2026: 19% of buyers exposed to genAI in their decision process reported lower confidence in their final purchase because of unreliable AI output.
ZenABM 2026 LinkedIn Ads benchmark (n=161,256 ads): Thought Leader Ads (promoted personal posts) deliver 2.68% median CTR versus 0.42% for single image ads, a roughly 6x lift.
Instantly 2026 Cold Email Benchmark: industry average reply rate is 3.43%, top quartile 5.5%, elite senders above 10%, with the gap driven by trigger-based personalization and human-written voice.
AI-as-editor is far less reputationally damaging than AI-as-author. Factual content (specs, pricing, comparisons) is largely insulated. Emotional, brand and founder content is the exposed surface.
Google's E-E-A-T guidance and the March 2024 and March 2026 core updates have rewarded named authors, original research and demonstrated experience while de-ranking thin, scaled, undifferentiated AI content.

Executive implications

Treat the brand voice guide as the prompt your AI will never read but your editors will. Ship a public, versioned guide signed by the CMO within ninety days.
Adopt a 30 / 40 / 20 / 10 production split: 30% strategy and brief, 40% humanization and edit, 20% AI generation, 10% QA and audit. The 40% humanization tranche is where most teams underinvest.
Run a 10-point pre-publish AI tell audit on every draft (banned-word check, structural ticks, em-dash density, burstiness, opener and closer formula, claim specificity, named human, real number, point of view, byline accountability).
Reallocate paid-social spend toward Thought Leader Ads on LinkedIn. The 6x CTR lift over single image ads is now the single highest-leverage paid-format change available to most B2B marketing teams.
Cap AI authorship in cold outbound at zero. Trigger-based, human-written outreach plus voice video (Loom, Vidyard) is the only channel-level move that closes the gap between the 3.43% industry average and the 10%+ elite reply rate.
Ban AI as final author on CEO posts, founder LinkedIn, all-hands emails and any internal communication. AI-as-editor is acceptable. AI-as-author is not.
Audit AI usage against EU AI Act Article 50 (enforcement begins 2 August 2026), China's September 2025 mandatory-labeling rules and the FTC final rule on AI-generated reviews. Plan for disclosure as a feature, not a fear.

Limitations

The Kirk and Givi finding is the strongest peer-reviewed evidence to date but is bounded to emotional marketing communications and consumer settings; B2B-specific replications are still limited. The Originality.ai vocabulary multipliers are derived from a single corpus and will drift as model providers tune output to suppress known tells. Burstiness signals are model-dependent and have already begun to compress as Claude and Gemini close the gap. Channel-level benchmarks (ZenABM LinkedIn, Instantly cold email) reflect platform medians and will vary by industry, deal size and audience. The case studies (Mailchimp, Linear, The Hustle, Anthropic) are illustrative; sustained discipline is harder to verify than published artifacts.

References

Kirk, C. P. and Givi, J. (2025). When AI authorship is detected, customers respond with moral disgust: seven preregistered experiments. Journal of Business Research, Vol. 186, article 114983.
TrustRadius. Bridging the Trust Gap: 2025 B2B Buyer Report. n=2,058 verified buyers, 490 vendors.
Reddit and SurveyMonkey. The Future of Brand Trust in the Age of AI. 2025, n=1,200.
Forrester. 2026 B2B Predictions. December 2025.
Edelman. 2025 Trust Barometer Special Report on Brands: From We to Me.
Originality.ai. AI Vocabulary Frequency Analysis. 3.3 million document corpus, 2024-2025.
Aityan, S. K. et al. (2024). StyloAI: stylometric detection of large language model output. arXiv preprint.
ProofreaderPro.ai (2025). Burstiness as a discriminator: 200 academic sample analysis.
Goedecke, S. (2025). Token frequency analysis of GPT-4o output. Personal blog and arXiv preprint.
The Last Fingerprint: How Markdown Training Shapes LLM Prose. arXiv, March 2025.
Lily Ray (2025). The AI Slop Loop: how Perplexity cited fabricated SEO articles about a Google update that never happened.
ZenABM. 2026 LinkedIn Ads Benchmark Report. n=161,256 ads across 211 companies.
Instantly. 2026 Cold Email Benchmark Report. Billions of interactions, 700,000+ businesses.
Averi. 2026 study on AI-generated comment engagement on LinkedIn.
La Growth Machine. 2026 LinkedIn content format benchmarks.
Vidyard. Personalized video prospecting platform benchmarks. 2025-2026.
Boords (2025). Video script template and spoken-versus-written language guidance.
Mailchimp. Public content style guide. styleguide.mailchimp.com.
Linear. Product changelog. linear.app/changelog.
Wiebe, J. (CopyHackers). Voice-of-customer rehab clinic case study (400% click lift).
European Commission. Artificial Intelligence Act, Article 50 enforcement timeline.
U.S. Federal Trade Commission. Final rule on AI-generated reviews, 2024.
Cyberspace Administration of China. Mandatory labeling regulation for AI-generated content. September 2025.

Key takeaways

AI slop has a detection-grade fingerprint. Delve appears 48 times more often in AI text than in human writing.
Burstiness collapses from a human standard deviation of 8.2 words per sentence to GPT-4o's 4.1. Readers feel the difference even when they cannot articulate it.
Kirk and Givi (JBR Vol. 186, 2025) measured a serial moral-disgust response to AI-authored emotional marketing across seven preregistered experiments.
73% of B2B buyers trust peer recommendations over AI chatbots. 62% of frequent AI users always or very often fact-check what they read.
AI-as-editor is far less damaging than AI-as-author. Factual content survives disclosure. Founder stories may not.

Decision brief

Generic AI language is not merely a style problem. In high-consideration B2B communication it signals low effort, weak audience knowledge, and interchangeable thinking. Trust rises when language contains a specific observation, a real constraint, and a consequence the buyer recognizes.

Decision criteria

The message could not be sent unchanged to another account.
Every claim is supported by context the recipient can verify.
The writing sounds like a person making a judgment, not a model completing a pattern.

Evidence to check

Reply quality and meeting conversion, not open rate alone.
Repeated phrases and sentence structures across outbound messages.
Buyer feedback on relevance, credibility, and specificity.

What I would do Monday

Review ten recent outbound messages. Delete every sentence that could apply to any company, then rebuild each opening around one verified business signal and one useful point of view.

Questions executives ask

Can B2B buyers really detect AI-generated copy?: Yes, and the evidence is no longer anecdotal. TrustRadius (2025, n=2,058 verified buyers) reports that 62% of frequent AI users always or very often fact-check the marketing content they read. Forrester's 2026 B2B Predictions show 19% of buyers exposed to genAI in their decision process felt less confident about their final purchase because of unreliable AI output. The pattern-recognition curve in cold outbound is now under three seconds. Buyers do not need to articulate why a piece of copy reads as AI. They down-weight the source and move on.
Is AI-as-editor the same risk as AI-as-author?: No, and the distinction is the most important operational rule in this guide. Kirk and Givi (JBR Vol. 186, 2025) explicitly carve out that the trust loss is attenuated when AI only edits a communication rather than authors it. The defensible workflow is to write the first draft as a human, use AI to tighten, condense and pressure-test, and then ship under a named byline. The damaging workflow is to ask the model for a draft, lightly skim, and publish.
Does this apply to factual product copy and documentation?: Largely no. The Kirk and Givi finding is also explicitly attenuated for factual messages, including pricing pages, technical specifications and comparison tables. Spec sheets, changelogs and API documentation can be AI-assisted heavily without measurable trust loss. The exposed surface is emotional and brand content: founder posts, About pages, customer stories, cold outbound, newsletters and LinkedIn.
What is the single highest-leverage edit a marketing team can make this week?: Build a banned-word file inside the brand voice guide, add the Originality.ai short list (delve, tapestry, multifaceted, leverage, robust, seamless, harness, unlock, in today's fast-paced world, it's worth noting), and forbid them in any AI-assisted draft without explicit override. Co-occurrence is the signal. Three banned words inside a 500-word passage raises AI-detection probability by 35 to 50 percentage points across modern classifiers. The cost of this change is one document and one editorial rule.
How do I audit my team's existing content without rewriting it all?: Run the 10-point AI tell audit against a representative sample of the last quarter's published content. Score each piece on banned-word density, structural ticks (it's not just X, it's Y), em-dash density, burstiness, opener and closer formula, claim specificity, named human presence, real numbers, stated point of view and byline accountability. Anything scoring under 6 out of 10 goes into a rewrite queue. Anything under 3 should be unpublished or noindexed.
Which channels demand the most human attention?: In order: cold outbound, founder and CEO LinkedIn, customer stories, About pages and newsletters. These are the surfaces where AI-as-author causes the most measurable trust loss. The channels where AI-as-editor is safest are technical documentation, changelogs, pricing pages, transactional emails and paid-ad creative variants.
Does Google actually penalize AI content?: Google penalizes AI slop, not AI assistance. The 2023 Search Central guidance, restated through the March 2024 and March 2026 core updates, rewards original, high-quality content that demonstrates E-E-A-T (expertise, experience, authoritativeness, trustworthiness) and de-ranks thin, scaled, undifferentiated content regardless of whether it was written by a human or a model. Sites with named authors, original research and demonstrated experience gained ground in both updates. Sites publishing programmatic AI SEO content lost majorities of indexed pages.
What KPIs prove the trust gap is closing?: Five leading indicators. Branded search volume on the company name and the founder name (rising). Reply rates on cold outbound (rising toward and through the 5.5% top-quartile threshold). Comment-to-impression ratio on LinkedIn personal posts (rising). Inbound demo requests citing a specific article or post (rising). And, on the lagging side, win rate against incumbents in deals where the buyer cited reading something the marketing team published (rising).

Canonical URL: https://juanbeltran.ch/blog/the-word-delve-is-costing-you-deals