Do AI-referred leads convert better than organic search leads

You cannot know whether AI-referred leads convert better than organic search leads until you tag those two populations separately and compare close rate on the same definition of a qualified opportunity. ChatGPT, Perplexity, and Google AI Overviews do not hand you a clean channel. Referrers get stripped, UTMs only ride URLs you control, and Overview clicks still arrive as google.com. Any slide that says “AI leads close 2x” without those holes is a marketing claim, not a measurement.

This page is the CRM split. Inclusion — whether ChatGPT or an Overview cites you — is a different scoreboard, and it lives in the Answer Engine Optimization playbook. I have been SEO certified since 2021. The AEO version of that work still ends in a named source field on the opportunity, not a sparkline of sessions.

A session is not a lead. A form fill is not an opportunity. Close rate is won ÷ (won + lost) inside a frozen cohort. If you mix those units, you will invent a lift.

The short answer

• Define the unit first. Opportunity close rate, same qualification bar, same sales motion. Not session-to-form. Not MQL volume.
• Tag a floor, not a census. Hostname allow-list for ChatGPT and Perplexity, plus UTMs on links you actually control. Copy-paste and in-app clicks will miss the net.
• Keep AI Overviews inside organic. They refer as Google. Do not “fix” that with a regex on google.com.
• Stamp analytics onto CRM. First-touch source, last-touch source, and a self-reported “how did you hear about us” field. If the three disagree, you do not have one AI close rate.
• Publish fractions. 8/20 versus 60/200. A percentage with no denominator is how a 2x rumor starts.

Population	Honest inclusion rule	Never include
AI-referred (tagged)	Session source host in the allow-list, or utm_medium=ai you set, or CRM original source written from that session	google.com, Direct, branded organic, “they mentioned ChatGPT in the demo”
Organic search	Google/Bing/DuckDuckGo organic, including Overview clicks you cannot split	ChatGPT/Perplexity hostnames, paid search, your own UTM’d newsletter
Unattributed	Direct, (not set), self-report only, assisted-but-not-first	Folding this pile into AI to make the sample bigger
Out of scope	Paid, partner referral, outbound	Mixing them into either closer

If you cannot fill that table for your stack this week, you are not ready to compare close rates. You are ready to instrument.

What counts as an AI-referred lead versus organic search?

An AI-referred lead is a person who reached a tracked URL from an answer product and became an opportunity you would have created anyway. ChatGPT Search, a Perplexity citation click, Claude with web tools, Gemini’s app — those are answer products. A Google AI Overview is also an answer product. It does not become an AI-referred lead in analytics, because the click still looks like organic Google.

Organic search is a click from a search engine results page (classic blue link or an Overview supporting link) with search as the source. That bucket is now mixed. Overview-informed buyers sit inside it. You cannot unmix them with document.referrer.

Label you want	What actually arrived	Put them in
ChatGPT citation click (desktop web)	Referrer host chatgpt.com or chat.openai.com	AI-referred, tagged
ChatGPT iOS/Android app click	Often no Referer	Unattributed until self-report or a UTM you control
Perplexity citation click (desktop web)	perplexity.ai or www.perplexity.ai	AI-referred, tagged
Perplexity app / in-app browser	Often stripped	Unattributed
Google AI Overview or AI Mode click	google.com / google.com/search, same as a blue link	Organic search (contaminated, documented)
User copied your URL out of the answer	Typed or pasted; Direct	Unattributed
User asked ChatGPT, then Googled your brand	Last click: branded organic	Organic last-touch; AI may be an assist, not the closer

Decision list — if you cannot check these, you do not have two populations:

• Written definition of “qualified opportunity” that sales already uses
• Written hostname allow-list that does not contain google.com
• Written rule for Overview clicks: stay in organic, footnote the mix
• Written rule for Direct: never auto-promote into AI
• Same close-rate formula for both buckets

ChatGPT is not “AI search.” Organic Google is not “not AI.” The comparison only works if the labels match the evidence you actually captured.

Why is a vendor “AI leads close 2x” claim usually unusable?

Because it almost never publishes the unit, the holes, or the denominator. Session conversion is not close rate. A demo booked from a ChatGPT click is not a closed-won. A sample of twelve AI-tagged forms is not a channel. And the tagged slice is a lower bound of true AI-influenced demand, so any “AI is hotter” story is running on the people who arrived with a referrer still attached — often desktop web, often already high-intent.

I will not put a Spurlock Studios lift percentage on this page. I have not measured a single blended “AI versus organic close rate” that survives the holes below. Inventing one would be a lie.

Claim you were shown	What is usually missing	What to demand
“AI traffic converts 2x organic”	Session-to-form vs opp-to-won mixed	Same unit, both sides
“ChatGPT visitors are more qualified”	Time-on-page, not won/lost	Closed-won ÷ closed-lost
“Our AI channel is tiny but elite”	Tagged floor treated as a census	Floor vs unattributed footnote
“Organic is dying, AI is the closer”	Overview clicks left inside organic	Split or a documented mix
A screenshot of GA4 AI Assistant	No CRM join	Opportunity IDs in the export

If the deck cannot show (a) the allow-list, (b) how CRM was stamped, and (c) won/lost counts, the 2x is decoration. Close the laptop.

Illustrative math (not a client result): 8 wins on 20 AI-tagged opps looks like 40%. 60 wins on 200 organic looks like 30%. One extra AI win moves the AI rate five points. You do not publish “AI converts better.” You publish 8/20 vs 60/200 and you wait.

What do UTMs actually capture, and where do they fail?

UTMs are query parameters you put on a URL. Google Analytics will prefer them over the Referer header when they are present (manual tagging vs auto-tagging). They do not appear because ChatGPT “ought to tag you.” If the model cites https://example.com/pricing with no query string, you get whatever referrer survived — or Direct.

Mechanism	What it tags	Failure
utm_source / utm_medium / utm_campaign on a URL you published	Clicks on that exact URL	Models often cite the canonical, untagged URL
UTM on llms.txt, partner roundups, or a tracked redirect you own	The subset of answers that reuse your tagged string	Many tools strip query strings before citing
UTM on every canonical (the temptation)	Everything that hits those URLs	You steal Google organic, email, and ads into “chatgpt”
No UTM, Referer present	Hostname only	Apps and no-referrer policies go dark
No UTM, Referer stripped	Nothing	Direct / (none) — see GA4’s (direct)/(none) note

UTM scheme if you tag anything you control

Keep it boring. One medium. Named sources. Never put this on the canonical that Google is supposed to rank.

1. utm_medium=ai
2. utm_source=chatgpt or perplexity or claude or gemini — match the product, not a mood
3. utm_campaign=citation or a page-level slug you can join to the URL
4. Publish the tagged URL only on surfaces you own (a tracked short link, a partner blurb, a PDF). Do not replace the canonical.

• Canonical URLs in Search Console remain untagged
• Tagged URLs 301 or serve the same content without creating a second indexable variant you did not mean to create
• CRM hidden fields capture utm_source, utm_medium, and landing path on submit
• Paid and email keep their own mediums; ai is not a junk drawer

UTMs are a flashlight you aim. They are not a census of answer-engine demand.

What does ChatGPT send as a referrer, and what does it swallow?

ChatGPT is several products sharing a brand. Desktop web citation clicks often arrive with host chatgpt.com, sometimes the older chat.openai.com. Native iOS and Android apps frequently send no Referer. In-app browsers and some share-sheet hops do the same. GA4 treats missing source as Direct / (none). That is not a ChatGPT bug you can ticket. It is how browsers and WebViews work when the Referer is omitted.

On May 13, 2026, Google Analytics added an AI Assistant default channel that assigns medium ai-assistant and campaign (ai-assistant) when the referrer matches a recognized assistant. The release note names ChatGPT, Gemini, and Claude. It does not publish the full host list. It does not claim it catches mobile apps that stripped the header.

ChatGPT surface (as of 2026)	Typical evidence on your site	Honest bucket
chatgpt.com citation click, desktop browser	Source host chatgpt.com	AI-referred, tagged
chat.openai.com leftover	Same idea, older host	AI-referred, tagged
ChatGPT iOS/Android app	Often Direct	Unattributed
User copied the URL from the answer	Direct or later branded search	Unattributed / last-touch organic
User asked ChatGPT, then clicked a Google result	Organic Google	Organic (AI may have assisted)

Do not invent a “ChatGPT capture rate.” Reports vary by site, app mix, and year. Log what you see in your host report for 28 days. That is your floor.

ChatGPT holes, in one list:

1. No Referer from apps. Floor undercounts mobile.
2. Copy-paste. The buyer never clicked the citation.
3. Brand search after the answer. Last-click organic eats the assist.
4. Unrecognized host. A new OpenAI hostname lands in Referral until you add it.
5. Native AI Assistant channel is not your CRM. Sessions without a lead join never enter the close-rate table.

If leadership only looks at chatgpt.com sessions, they are looking at desktop leftovers. Say that out loud.

Verify the ChatGPT hole on your own domain

Do not argue from a blog post. Click a citation of your site from the products your buyers name.

1. Open ChatGPT Search (desktop browser, logged in the way a buyer would). Ask a prompt that already cites you, or one that should.
2. Click your URL. Confirm the landing URL is the canonical (no surprise query string).
3. In GA4 DebugView or a realtime report, read Session source / medium within a minute.
4. Repeat from the ChatGPT iOS or Android app if that is a real buyer surface. Expect Direct. If you get a host, add it.
5. Repeat with a copied URL pasted into a new tab. Expect Direct.

Result	Meaning	Sheet rule
chatgpt.com / ai-assistant or Referral	Desktop web floor works	Count as AI-referred
(direct) / (none) from the app	App stripped Referer	Unattributed; self-report may catch it
google / organic after you Googled the brand	Last-click ate the assist	Organic last-touch; do not refile as AI
No session at all	Tag blocked, consent, or wrong property	Fix collection before you compare close rate

A missing DebugView hit is a collection bug, not proof that “ChatGPT users never convert.”

What does Perplexity send, and what still lands as Direct?

Perplexity is the answer product that most often does send a Referer on desktop web citation clicks — perplexity.ai and www.perplexity.ai both show up in real properties. Google’s June 11, 2026 Analytics update even calls out built-in Source Group grouping for ChatGPT (OpenAI) and Perplexity. That is source hygiene. It is not close rate.

The May 13 AI Assistant note named ChatGPT, Gemini, and Claude. It did not name Perplexity. Do not assume the default AI Assistant channel caught Perplexity on day one. Put perplexity.ai and www.perplexity.ai in your custom channel so you are not waiting on Google’s private list.

Perplexity path	Referrer you should expect	If you omit www
Desktop web citation	perplexity.ai or www.perplexity.ai	www falls through to Referral
Mobile app / in-app	Often none	Direct
Shared Perplexity thread opened in a browser	Host may be Perplexity — or Direct after a redirect	Test the actual share URL
Copied source URL	Direct	Unattributed

Hostname allow-list (start here, then add what your logs show)

1. Export Session source for 90 days.
2. Filter to hosts that are obviously answer products.
3. Add only those hosts. Do not add google.com.
4. Re-export 14 days later. New hosts appear. Update the regex.

Conservative regex for a custom channel — not Google’s help-center example:

^(chatgpt\.com|chat\.openai\.com|perplexity\.ai|www\.perplexity\.ai|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com)$

Google’s custom channel example uses a sketch regex that matches fragments like ai, gpt, and bard with wildcards. That pattern will eat unrelated sources (anything containing ai is a live footgun). Use anchored hostnames. Place the channel above Referral. Do not place it above Organic Search in a way that could steal Google.

Perplexity is the easy screenshot. ChatGPT apps and Overviews are where the comparison dies. Instrument Perplexity. Do not let it become the whole story.

Why do AI Overview clicks look like Google organic?

Because they are Google Search. The click leaves google.com (or a country variant). There is no public Referer token that says “this supporting link sat inside an Overview.” Search Console’s generative AI performance report tells you impressions for AI Overviews and AI Mode — pages, countries, devices, dates. As of the help page, that dedicated report is still an inclusion lens, not a click ledger you can join to CRM. Google launched it June 3, 2026; as of August 31, 2026 they say it rolled out worldwide. Impressions are not opportunities.

If Overviews ate informational clicks, that is a page-job problem — treat it in If AI Overviews ate your clicks. It is not a close-rate split you can read from GA4 Organic Search.

Signal	What it proves	What it does not prove
GSC gen-AI impressions on /pricing	A link to that URL was shown in Overviews or AI Mode	Someone clicked, became a lead, or closed
GA4 google / organic on /pricing	A Google Search click landed	Overview vs blue link
#:~:text= fragments on some Overview clicks	Directional, vendor-dependent, easy to confuse with snippets	A complete AIO channel
Branded-query lift after an Overview week	Possible view-through	Not a tagged AI-referred lead
Self-report “I saw you in Google’s AI summary”	First-party, messy, useful	Not automatically an Overview click

Rules you write down:

1. Never classify google.com as AI-referred.
2. Footnote organic close rate: “includes Overview clicks we cannot split.”
3. Do not “correct” organic by subtracting GSC gen-AI impressions. Impressions are not clicks.
4. If you need Overview-specific outcomes, use a survey field or a sales note — and keep that as a third column, not a merge.

Schema will not save this. JSON-LD that answer engines can use may help a system cite the right entity. It does not mint a utm_source=ai-overview on the click. Markup is not a conversion tag.

How do you build an analytics channel that is a floor, not a ceiling?

You build a custom channel group so the rule is yours, and you keep the native AI Assistant row as a cross-check after May 13, 2026. Custom channel groups apply retroactively in reports. The Default Channel Group rewrite for AI Assistant is a forward classification of recognized referrers. If you need history, the custom group is the one you can defend.

Limits: standard properties get 2 custom groups besides the predefined one; 360 properties get 5. Each group maxes at 50 channels. Do not burn a slot on a vanity “GPT vs Claude” split until the joined CRM table exists.

Procedure — GA4, one sitting

1. Admin → Data display → Channel groups → Create new (copy Default).
2. Add channel AI assistants (allow-list) with Session source matches regex (anchored hosts above).
3. Optionally add a second channel AI assistants (UTM) for medium exactly ai.
4. Reorder: UTM AI, then hostname AI, then Organic Search, then Referral. Confirm google still hits Organic Search.
5. Save. In Traffic acquisition, set this group as the primary dimension. Spot-check 14 days of chatgpt.com and perplexity.ai.
6. Export Session source / medium. Anything new and obviously an assistant gets added next week — not in a panic on day one.

Check	Pass	Fail
chatgpt.com	Your AI channel	Still Referral
www.perplexity.ai	Your AI channel	You forgot www
google / organic	Organic Search	You put google in the AI regex
medium=ai on a tagged URL	UTM AI channel	Canonical accidentally tagged
(direct) / (none)	Direct	You recoded Direct as AI

• Channel sits above Referral
• google.com is not in the regex
• Direct is untouched
• BigQuery or a weekly CSV export exists so CRM can join landing page + time
• Someone owns the regex when OpenAI ships a new host

The ceiling is “every buyer who used an answer product.” You will never see that number in GA4. The floor is the tagged set. Close rate uses the floor, with the ceiling named as unknown.

Layer	What Google actually shipped	Use it for	Do not use it for
Default Channel Group AI Assistant (May 13, 2026)	Medium ai-assistant, campaign (ai-assistant) when the referrer matches Google’s recognized list (release note)	Forward-looking session slice for ChatGPT, Gemini, Claude (named)	History, Perplexity-unless-confirmed, CRM close rate
Custom channel group	Your regex; retroactive in reports	Defensible floor you can screenshot	A regex so wide it eats email / paid
Source Group (June 11, 2026)	Built-in grouping that names ChatGPT (OpenAI) and Perplexity among others	Cleaning messy source strings	Opportunity won/lost
Edge / CDN logs	Raw Referer even when a tag is blocked	Catching new hosts	Calling a log line a lead

Standard properties: 2 custom groups, 50 channels each. Spend the slot on the allow-list, not on a vanity GPT-vs-Claude split, until CRM is joined.

If the native AI Assistant row and your custom channel disagree, export both for the same date range. The delta is usually www variants, a host Google has not listed, or UTMs you applied. Write the reason next to the number. Do not average the two rows into a “true AI traffic” figure.

How do you stamp the same split onto CRM leads?

Analytics without a CRM field is a session report. Close rate lives on opportunities. On form submit (and on sales-created records), write the acquisition evidence onto the record once, then stop letting last-click overwrite the original unless you are storing last-touch in a second field.

CRM field	Writes from	Role in the comparison
Original source	First session with a known source, or first form UTM/referrer	Primary AI vs organic split
Latest source	Session that converted	Last-touch check, not the only story
Landing page + timestamp	First hit	Join back to GA4
UTM source / medium / campaign	Hidden fields	Only if present
Referrer host	First-party capture on landing, if you collect it	Survives some CRM imports that drop GA4
Self-report	“How did you hear about us?”	Catches copy-paste; never auto-merge into Original
Qualification	Sales process, not marketing	Same bar both buckets

Procedure — form to opportunity

1. Hidden fields: utm_source, utm_medium, utm_campaign, landing_page, gclid (exclude paid from this study).
2. Visible field: How did you hear about us? Options include Google search, ChatGPT, Perplexity, another AI assistant, referral, other.
3. On create, map hostname or utm_medium=ai → Original source = AI-referred. Map google / organic (and Bing organic) → Organic search.
4. Create a second field for self-report. Do not overwrite Original source with the dropdown.
5. Opportunities inherit the contact’s original source. Sales may add an assist note. Notes are not a source rewrite.
6. Weekly: export opps created in the window with Original source, stage, close date, won/lost.

• Marketing cannot change Original source without an audit log
• Self-report is a separate column in the close-rate sheet
• Paid (gclid, cpc) is excluded from both closers
• Unattributed stays Unattributed
• Demo notes like “they use ChatGPT at work” do not refile the opp

If your CRM only has one source field, you will fight last-click forever. Add the second field before you brief a close-rate chart.

Joining analytics to CRM is a timestamp plus a landing page plus, if you have it, a first-party identifier. You do not need a warehouse on day one. You do need a rule for collisions.

Join key	Good enough when	Collision
Form submit time ± 30 minutes + landing path	Low volume, one form	Two sessions, one submit — keep Original from first known non-Direct
Email / User-ID present in both	Logged-in app or portal	Marketing site has no User-ID — do not fake one
Hidden UTMs only	Tagged URLs you own	Empty UTMs — fall back to referrer host
Sales-created opp, no form	Rep asked the source question	Self-report only; Original = Unattributed unless they came through the site

• Paid IDs (gclid) never map to AI-referred
• A second form fill does not overwrite Original source
• Offline opps without a site session stay Unattributed or self-report-only
• The join is documented in one paragraph a new hire can run

Self-report is a gift and a trap. “ChatGPT” on a dropdown catches copy-paste. It also catches people who use ChatGPT at work and found you on Google. Keep it beside Original source. If 40% of organic opps tick ChatGPT and 4% of sessions are chatgpt.com, you are measuring brand awareness of a tool, not citation clicks.

How do you compare close rate without inventing a lift percentage?

Pick one formula. Print it on the sheet. Do not switch it when the number looks sad.

Close rate = closed-won ÷ (closed-won + closed-lost) for opportunities that entered the cohort in a fixed created-date window, after the same qualification gate, excluding still-open deals from the denominator or aging them out with an explicit rule. Pick one. Write it down.

Do not use:

• Sessions with a form event ÷ sessions (that is a conversion rate, different job)
• Open pipeline as implicit losses
• MQLs that sales never accepted
• A blended “AI-influenced” pile that includes assists

Report this	Format	Ban
Cohort window	Created 2026-05-01 to 2026-07-01, closed by 2026-09-01	“This quarter, vibes”
Counts	AI tagged: 8 won, 12 lost (8/20). Organic: 60 won, 140 lost (60/200)	“AI is 40%, organic is 30%” as a headline with no counts
Still open	6 AI, 40 organic — listed, not stuffed into lost	Quietly dropping opens
Unattributed	4 won, 10 lost — third row	Merging into AI “because ChatGPT is big now”
Assists	Count of organic opps with ChatGPT in self-report	Adding them to the AI numerator

Procedure — the weekly closer sheet

1. Filter opportunities: created in the window, qualified, not paid, not partner.
2. Split by Original source: AI-referred, Organic search, Unattributed, Other.
3. For each split, count won, lost, still open.
4. Compute close rate only where (won+lost) is large enough that one additional win does not move the rate by a story-sized amount. If it does, print the fraction and stop.
5. Print last-touch and self-report as extra columns. If they disagree with Original source, say “split is unstable” instead of averaging them into a fake lift.
6. Do not compute “AI is Nx organic.” If you must compare, show both fractions on one line and the sample sizes.

I will not tell you AI-referred leads convert better. I will tell you how you would know. Until the sheet exists, the honest sentence is: we have not measured it.

When Original, last-touch, and self-report disagree, do not average them into a blended AI rate. Pick a primary (Original source) and show the others as sensitivity.

Original	Last-touch	Self-report	What you write
AI-referred	Organic branded	ChatGPT	AI first-touch closer; last-click is brand search after the answer
Organic	AI-referred	Google search	Last-click AI on an organic-origin opp — do not steal it into the AI closer if Original is the rule
Direct	Organic	Perplexity	Unattributed original; Perplexity is a hint, not a stamp
Organic	Organic	ChatGPT	Organic closer; ChatGPT is an assist or a tool habit — footnote, do not merge
AI-referred	AI-referred	Google search	Tagged AI closer; the dropdown is noise

Three reports, one sentence each, beats one “true” percentage. If sales enablement needs a single number, give them the Original-source fraction and the sample size, not a multiplier.

What sample size and time window keep the comparison honest?

A sales cycle is the clock, not your content calendar. If median days-to-close is 40, a two-week “AI vs organic” readout is a form-fill report wearing a closer costume. Age the cohort: created-date window, then a close-date cutoff at least one median cycle later. Still-open deals at cutoff are a third status, not losses.

There is no sacred n. There is a stability test: add one hypothetical win to the smaller bucket. If the percentage you wanted to tweet changes enough that the story flips, you do not have a story.

Situation	What to do	What not to do
AI tagged (won+lost) < 20	Print 8/15, no %	“AI converts at 53%”
Organic is 10× larger	That is expected; do not downsample organic to “be fair”	Truncating organic to match AI
Window shorter than median cycle	Report pipeline, not close rate	Early “AI is winning”
Mix of PLG self-serve and sales-led	Split motions first	One close rate across both
Seasonality / one huge deal	Call the outlier	Let one $200k win define AI

Checklist before anyone puts the comparison in a board deck:

• Median days-to-close known for this motion
• Created window ≥ one cycle; close cutoff ≥ one cycle after the last created date
• Same qualification gate both sides
• Outliers listed
• Unattributed shown
• Overview contamination footnoted on organic
• No Nx multiplier

If the tagged AI bucket is tiny, the correct investment is more instrumentation and more citation work in the AEO playbook — not a hot take on close rate. A floor of twelve opportunities is a queue, not a channel.

Aging calendar (illustrative, not a forecast)

Assume median days-to-close is 40. You turned tagging on July 1.

Date	What you may say	What you may not say
Jul 8	Channel exists; N sessions tagged	Close rate
Jul 31	N opps created with Original = AI-referred	“AI is closing better” (too many still open)
Sep 9	Created Jul 1–31, closed by Sep 9: print won/lost/open	Treating remaining open as lost to juice organic
Oct 1	Second month created, same cutoff rule	Changing the formula because month one looked boring

Worked stability check (illustrative counts, not a client): AI tagged 7/18 (39%). Organic 55/190 (29%). Add one AI win → 8/19 (42%). That jump is larger than the organic gap you wanted to celebrate. You still do not have a headline. You have a queue that needs more closed deals.

If two motions share a CRM (product-led signup vs sales-led demo), split before you compare. A ChatGPT click that starts a self-serve trial is not the same closer as a six-week enterprise demo that started from a non-branded Google query. Mixing them manufactures a lift from mix shift.

What fails first when teams try this?

The comparison dies in the contaminated organic bucket and the unit mismatch. Those two beat “we forgot Perplexity” every time.

Failure mode 1 — unit mismatch. Marketing charts GA4 session conversion for the AI Assistant channel against Salesforce close rate for organic opportunities. AI looks like a miracle or a corpse depending on which way the units lean. Cost: a quarter of content budget pointed at the wrong surface. Fix: one object (qualified opportunity), one formula, two source labels.

Failure mode 2 — organic eats Overviews, then loses the narrative. Overview-informed buyers are mixed into organic. Someone reads a traffic drop on informational URLs (the Overview click-job post) and declares organic “low quality.” You cannot prove that from mixed google.com clicks. Footnote the mix or stop comparing.

Failure mode 3 — UTM on the canonical. A well-meaning ?utm_source=chatgpt on URLs listed in llms.txt gets copied into Google results, Slack, and the homepage. Organic, Direct, and email start reporting as ChatGPT. The AI closer goes to 100% because you poisoned the well. Fix: never UTM the canonical.

Failure mode 4 — Direct recoded as AI. “Direct spiked on URLs we know are cited, so Direct is AI.” That guess will also eat bookmarks, Slack, and iOS. Keep Direct in Unattributed. Use self-report if you need a hint.

Breakage	What it costs	Instead
Session CR vs opp close rate	Fake 2x	One formula
google.com in the AI regex	Stolen organic	Allow-list without Google
Canonical UTMs	Permanent source lie	Tagged redirects you own, only
Direct → AI	Inflated AI, angry sales	Self-report column
Sales rewrite of Original source	Un-auditable chart	Assist notes, second field
n=11 treated as a rate	A LinkedIn carousel	Fraction only

• Canonical URLs have no AI UTMs (spot-check GSC)
• Close-rate sheet uses opportunities, not sessions
• Organic footnote mentions Overviews
• Direct is not in the AI numerator
• One person can explain the allow-list without opening Notion

The expensive version of this failure is a keynote. The cheap version is a sheet with three rows and no multiplier. Pick cheap.

What should you skip if you only have a week?

You skip the verdict. You ship the floor. A week is enough to stop lying. It is not enough to know whether AI-referred leads close better.

1. Day 1: Write the opportunity definition and the three-bucket source list (AI tagged, organic, unattributed). Ban google.com from AI.
2. Day 2: Custom channel in GA4 with anchored hosts. Export last 28 days of Session source. Add www.perplexity.ai if it appears.
3. Day 3: Hidden UTM fields + “how did you hear about us” on the form. Do not UTM canonicals.
4. Day 4: CRM Original source mapping. Second field for self-report. Stop sales from overwriting Original.
5. Day 5: Sheet with won/lost/open counts. If (won+lost) is tiny, you publish counts and a date when the cohort ages out. No lift.

This week	Next month	Not this quarter
Allow-list channel	Join to CRM at opp create	Multi-touch model as the only slide
Form fields	Aged close-rate cohort	“AI closes 2x” in sales enablement
Self-report options	Stability test (one extra win)	Recoding Direct
Footnote on Overviews	Review new hosts	Subtracting GSC impressions from organic

Skip vendor AI-visibility scores as a closer proxy. Skip a prompt panel as a substitute for source fields — that panel measures inclusion, not close rate. Skip schema as measurement. Skip renaming Organic Search to “legacy.”

A week of instrumentation beats a quarter of anecdotes. The anecdote is how the 2x rumor travels.

When is this comparison not worth running yet?

When you cannot name a qualified opportunity, when you have no form or CRM, when tagged AI opportunities are a handful, or when sales and marketing do not share a source field. Also when the site’s conversion path is broken — quiet phones and dead forms are a different diagnostic. Fix the path before you argue about channel quality.

Gate	If missing	Do this instead
Shared opp definition	You will compare two products	One-page definition, signed by sales
Working form/CRM	Nothing to stamp	Capture before charts
Allow-list in analytics	You will use Referral folklore	Hostname channel first
Enough closed opps to survive +1 win	You will invent a %	Keep counting
Organic is 90% branded	“Organic vs AI” is brand vs citation	Split branded vs non-branded organic first
You sell one $400k deal a year	n is the deal, not the channel	Narrative + source notes, no rate

• You would bet a hiring plan on this split today — if no, do not present a rate
• You can show the allow-list and the CRM map in one screenshot
• You can say “we don’t know” about Overviews without flinching
• You are willing to publish fractions with small n

Not worth it yet is a legitimate AEO outcome. Get cited, get tagged, then compare. Reverse that order and you will hire a narrative.

FAQ

Do AI-referred leads convert better than organic search leads?

Not as a known fact. They convert better only if your tagged AI-referred opportunities close at a higher rate than organic opportunities under the same qualification bar, and you still have to footnote that the AI tag is a lower bound and that Overviews sit inside organic. I will not invent a lift percentage. Build the split, then read the fractions.

How do I measure whether AI-referred leads convert better than organic search leads?

Tag a hostname-and-UTM floor in analytics, stamp Original source on the CRM opportunity, age the cohort through your sales cycle, then compute won ÷ (won + lost) for each bucket. Publish counts. Keep self-report and last-touch as extra columns. If one extra win flips the story, you are still counting, not comparing.

What usually fails first when teams try this?

Unit mismatch and a contaminated organic bucket. Teams compare GA4 session conversion to CRM close rate, or they dump AI Overview clicks into a homemade AI channel, or they UTM the canonical and poison every other source. Direct recoded as AI is the runner-up. Fix the object and the allow-list before you brief a rate.

How long does this take to show results?

Instrumentation can stand up in a week. A close-rate comparison takes at least one full sales cycle after tagging starts, plus a close-date cutoff. Mid-cycle “AI is winning” is a pipeline anecdote. If your median close is 45 days, plan on a created window plus those 45 days before you argue with the organic number.

What should I skip if I only have a week?

Skip the verdict, the Nx multiplier, recoding Direct, and any UTM on canonical URLs. Ship the allow-list channel, the form fields, the CRM Original source map, and a sheet that only prints counts. Inclusion work (citations, schema, Overview page jobs) can continue in parallel; it is not this comparison.

When is this not worth doing yet?

When you lack a shared opportunity definition, a working CRM stamp, or enough closed deals in the tagged AI bucket to survive a one-win stability test. Also when the conversion path itself is broken, or when “organic” is almost all branded navigational search. Instrument first. Compare later. “We have not measured it” is an acceptable board sentence.

CTA

You cannot compare AI-referred close rate to organic until the two populations are tagged. Build the floor, stamp CRM, then report fractions — not a lift slide.

Lane: /visibility · Book a visibility audit.