What is share of voice in AI answers, and how do I measure it

Share of voice in AI answers is your slice of the named brands and credited sources on a frozen buyer-prompt panel, versus a frozen competitor set, computed separately for each engine. It is not classic SEO share of voice, and it is not a Search Console impression sparkline. Measure it as two ratios — mention SOV and citation SOV — on the same prompt IDs every week. If you cannot name the prompts, the competitors, and the engines, you do not have share of voice. You have a screenshot.

This spoke sits under the Answer Engine Optimization playbook. The inclusion definition lives in what AI visibility is. Lane work lives on /visibility. I have been SEO certified since 2021. Certification does not invent a category SOV target. The sheet does.

The short answer

• Freeze 25–40 buyer prompts, 3–6 competitors, and a scoring codebook before you change a page.
• Score every run as mention, citation, both, or neither. Recommendation is a third flag on money prompts, not a replacement for the first two.
• Compute mention SOV and citation SOV per engine. Never average ChatGPT, Perplexity, and Google AI Overviews into one “AI SOV %.”
• Semrush’s published formula is mentions ÷ all tracked-brand mentions. Their Enterprise product also weights position and, for ChatGPT, topic volume. Label which formula you used.
• I will not quote an industry “good SOV.” Vendor index studies describe their prompt mix. Your frozen panel is the only denominator that belongs on your slide.

What is share of voice in AI answers?

Share of voice here is a competitive ratio inside generated answers: how often you appear relative to the other brands you actually lose deals to, on questions a buyer would ask. Semrush’s knowledge base defines Share of Voice as the percentage of mentions your brand receives in AI-generated answers compared to competitors in your market. That is the right object. It is not a rank, a session, or a vendor grade you cannot audit.

Object	What it measures	What it needs	What it is not
Classic SEO SOV	Weighted presence in a keyword SERP set	Rank tracker + keyword list	Inclusion in a generated answer
Paid / social SOV	Spend or volume versus rivals in a channel	Ad or social dataset	ChatGPT or Overview presence
Mention SOV (AI)	Your brand-name events ÷ tracked-brand mention events	Frozen prompts + frozen competitors	Proof you were the source
Citation SOV (AI)	Your credited-URL events ÷ tracked-domain citation events	The same panel, a visible source chip or URL	A guarantee anyone clicked
GSC gen-AI impressions	How often your links were shown in Google AI features	Search Console	Any competitor denominator

Operator test: if the number still exists after you delete the competitor column, it is not share of voice. It is a rate about you.

Why is classic SEO share of voice the wrong object?

Classic SOV answers “what share of the blue-link real estate did we occupy for this keyword set?” Answer products do not sell real estate that way. One Overview is a single SERP slot that several URLs share. A ChatGPT answer can name five brands and cite fifteen URLs. A Perplexity answer can footnote a Reddit thread you do not control. Mapping last year’s keyword SOV onto that object is how teams celebrate a #1 they were never named in.

Classic SEO SOV assumes	AI answers actually do	Reporting damage if you import the old number
One position per query	Several brands can share one answer	You under-count rivals who were co-named
Your URL either ranks or does not	You can be named with zero owned URL	You miss hollow mentions
Keyword list ≈ demand	Buyer grammar is prompt-shaped	You measure jargon nobody asks
Google is the market	Buyers split across chat and Search	You grade the wrong engine
Impression share is competitive	GSC gen-AI impressions are your links only	You label a first-party chart “SOV”

I still run competitive SERP tools the way I use a packing list: to see shape on classic results. They do not replace a prompt-panel denominator. If leadership wants one chart titled share of voice, it has to be the panel, or it has to say “classic SEO” in the title.

Checklist before you reuse an old SOV dashboard:

• Keyword list and prompt list are two tabs, not one mashed column
• Rank still reported for eligibility, never as AI inclusion
• Competitor set for AI is the brands sales actually loses to, not the domains that rank #2–#5
• Title of the slide names the engines
• No GSC impression line labeled SOV

Wrong object in, wrong tickets out.

What is a mention versus a citation on the same answer?

A mention is the brand string in the answer. A citation is visible credit to a URL, footnote, source chip, or explicit attribution. schema.org’s citation property is a reference to another creative work — a page, article, or publication — not a name in prose. Keep that split in the sheet even if a vendor blends it.

Semrush’s measure-AI-SoV guide (July 17, 2026) draws the same four-way split operators need: mention, citation, source, and share of voice. A source is a page the system retrieved. Retrieved-but-uncited is not a citation. Semrush’s 2026 AI Visibility Index (June 26, 2026; 126 million U.S. prompts, January–April 2026) reported that on Gemini the overlap between mentioned brands and cited domains can be as low as 30%. That is Semrush’s mix, not your vertical. It is still the reason you keep two SOV series.

State	Mention	Citation	What the buyer got	What you log
Named and credited	Yes	Yes	Name plus a path to you	mention, citation, URL
Named only	Yes	No	Awareness, no proof	mention only
Credited only	No	Yes	A URL without a brand shout	citation only — fix the lead
Recommended	Yes	Optional	“Use this one” on a money prompt	Extra recommended flag
Absent	No	No	The room without you	both no
Wrong	Either	Either	A false fact with your name on it	accuracy: fail — negative visibility

Operator test: delete every link and source chip. What still names you is a mention. What disappeared was the citation surface.

Do not average mention SOV and citation SOV into one “visibility %.” High mention / low citation funds brand-name blog posts. High citation / low mention funds entity work on the URL that already got the footnote.

How do you freeze a prompt panel SOV can survive?

A prompt panel is a versioned list of buyer questions you re-run on a calendar. SOV dies when the list, the competitors, or the scoring rules move mid-quarter. Ad hoc Slack screenshots are anecdotes. They cannot produce a ratio.

Panel rule	Why SOV needs it	What happens if you skip it
25–40 prompts	Enough rows to trend; few enough to actually run	200 prompts you never finish, or 8 vanity questions
Prompt IDs (p_014)	Wording changes are versions, not silent edits	You compare two different questions and call it a lift
Frozen competitor set (3–6)	The denominator is a decision	Adding a weak rival inflates you
Bucket tags	Recommendation ≠ how-to ≠ brand	Category SOV gets polluted by “who is [you]?”
Engine list	Each product is a market	One blended % hides the hole
Owner + revenue tag	Someone has to run it	The sheet rots in a week

Bucket the panel before the first run:

Bucket	Example shape	Role in SOV
Category / recommendation	“best X for a 20-person team”	Primary competitive SOV
Comparison	“A vs B for [constraint]”	Where deals die — weight this
How-to / problem	“how to [job] without [pain]”	Citeable; often low revenue — report separately
Local / ICP	city, stack, compliance, budget	Stops generic vanity wins
Brand / reputation	“[you] pricing,” “[you] vs [them]”	Accuracy series — not category SOV

Brand prompts belong on the panel. They do not belong in the category SOV denominator. “Who is Acme?” will name you. Mixing that row into competitive share of voice is how teams fake a win.

Procedure to freeze the panel in one afternoon:

1. Pull 15 questions from sales notes and lost-deal writeups, not from your blog titles.
2. Add 10 comparison and recommendation prompts a buyer would type without your brand.
3. Add 5 brand/reputation prompts and 5 local or ICP-flavored prompts.
4. Name 3–6 competitors sales will defend in a meeting. Lock the list for the quarter.
5. Write the scoring codebook (next section) on the same sheet. Then run. Do not wait for a taxonomy in Notion.

Minimum columns on the log. If a column is missing, you will invent it later under pressure.

Column	Type	Rule
prompt_id	p_014	Never reuse an ID for new wording — bump v2
bucket	enum	recommend / compare / howto / local / brand
engine	enum	One product per row
run_date	ISO date	Same calendar window for a cycle
mention_you	0/1	Visible brand string only
citation_you	0/1	Visible URL / footnote / chip only
citation_url	URL or blank	The credited page, not the homepage you wish they used
recommended_you	0/1	Money buckets only; blank elsewhere
accuracy	pass / fail / na	na if you were not discussed
rivals_named	slugs	Locked set only — pipe-separated
domains_cited	domains	Include Reddit, Wikipedia, directories as sources, not as rivals
notes	short	UI quirks: Search on/off, collapsed Overview, mobile

• Competitor tab lists legal names and the strings the model actually uses
• Brand-bucket rows can be filtered out of category SOV in one click
• Empty runs are not_run, never filled from memory
• Screenshots or exports live next to the week, not in a random Slack channel

Without those columns, share of voice becomes a vibe with a percentage sign.

Why does one blended SOV hide the engine that matters?

Citation pools barely overlap. The ChatGPT vs Perplexity vs AI Overviews spoke is the engine map; this spoke is why you must not average them. Semrush’s Index, on their four-platform mix, said ChatGPT cited an average of 15 sources per response while Gemini cited an average of 3. Different slot counts change what “share” even means. A 20% citation SOV on a 15-source answer is not the same object as 20% on a 3-source answer.

Engine	What you can observe	SOV you can honestly compute	Common cheat
ChatGPT Search	Name in the answer; Sources panel / cited URLs	Mention SOV + citation SOV on that UI	Counting training-data vibes with Search off
Perplexity	Inline numbers, domain footnotes	Same two series	Treating a Reddit cite of a rival as your citation
Google AI Overviews	Named in the unit; supporting links	Manual SERP log + citation SOV	Calling GSC impressions “Overview SOV”
Google AI Mode	Follow-up threads; links in the answer	Prompt you typed + visible credits	Merging Mode with Overviews because both are Google
Other chat UIs	Product-native cards	Only if buyers actually use it	Adding Copilot for a pie chart nobody asked for

Priority follows where your buyers already ask, then deal influence. Keep a thin baseline on the engines you are not funding. Instrumented neglect beats an unmeasured panic screenshot from a surface that does not close revenue.

Checklist per weekly cycle:

• Same prompt IDs on every in-scope engine
• Search-on noted for ChatGPT when the UI offers it
• Device class noted for Overviews (mobile hides links)
• Three SOV rows per engine: mention, citation, recommended (money bucket only)
• No cell named “overall AI SOV” unless it is clearly a secondary appendix

If you only have one number, pick the engine that influenced the last five deals. Do not pick the engine that looked nicest in a vendor pie.

What is the measurement recipe?

This is the whole program. Tools are optional. The ritual is not.

Ingredients (frozen before week 1)

• Prompt panel with IDs, buckets, owners, revenue tags
• Competitor set (3–6) locked for the quarter
• Engine list your buyers use
• Scoring codebook: mention, citation URL, recommended, accuracy, other brands named
• A sheet that can compute ratios from counts — not a slide that starts with a percentage

Recipe

1. Freeze. Version the prompt list (panel_v1). Do not silently rewrite p_014 when it makes you look absent.
2. Run. Same calendar window, same prompt order, each in-scope engine. Screenshot or export enough to defend a cell in a meeting.
3. Score mention vs citation. Binary on the prompt for you. List every tracked brand named. Paste cited domains into their own cells.
4. Split by engine. Copy the scoring block once per product. Do not collapse rows.
5. Compute two SOVs per engine. Mention SOV = your mention events ÷ sum of mention events for you + locked competitors. Citation SOV = your citation events ÷ sum of citation events for tracked domains. A run that names you and two rivals contributes three mention events, not one.
6. Keep brand-bucket rows out of category SOV. Report accuracy on brand prompts as a separate series.
7. Write the counts on the slide. “Named in 8 of 20 recommendation prompts on ChatGPT Search; Competitor A named in 11. Mention SOV 8 ÷ (8+11+6+3) = 29%.” If you cannot show that sentence, delete the percentage.
8. Ticket the gaps. Every material miss becomes a page, fact, schema, or corroboration ticket. A sparkline with no owner is theater.

Step	Artifact	Done looks like
Freeze	panel_v1 + competitor tab	Dates, owners, no “TBD rivals”
Run	Dated log, one row per prompt × engine	Empty cells marked not_run, never guessed
Score	Mention / citation / recommended / accuracy	Codebook applied, not vibes
Split	Engine sheets or a engine column	No blended default view
Compute	Two SOV columns per engine	Formula visible, not a vendor black box
Report	Counts + ratio + window	Competitor list on the same slide
Ticket	URL + owner + due date	Absence has a next action

Run weekly for money buckets if you can staff it. Biweekly is the floor. One run is an anecdote. Three runs is the start of a trend. Do not brief a board off a single Tuesday.

How do you score a run without collapsing the columns?

Write the codebook on the sheet. If two people would score the same answer differently, your SOV is a coin flip.

Code	When to use it	Do not use it when
mention_yes	Brand string or unambiguous nickname appears in the visible answer	The domain appears only in a hidden source list the user never saw
citation_yes	A URL, footnote, source chip, or “according to [your domain]” is visible	You believe the model “used” you. Belief is not a citation
recommended	The answer picks you (or a short list including you) as the thing to use, hire, or buy	You were named in a long roundup with no preference
accuracy_fail	Material error on price, city, offer, or who you are	A tone you dislike
competitor_[slug]	Tracked rival named	A random directory, Wikipedia, or Reddit as if they were a rival brand

Procedure for one row:

1. Paste the prompt ID and engine.
2. Read the visible answer. Mark mention_yes or mention_no.
3. Inspect source chips / footnotes / Overview links. Mark citation_yes only if your domain is credited in that UI.
4. If the prompt is a recommendation or comparison, mark recommended yes/no.
5. List every locked competitor named. Do not add a new competitor because they showed up once.
6. If the answer is about you, mark accuracy. Wrong plus named is not a SOV win.

Teaching sheet — not a client result, not an industry average. Twenty recommendation prompts, one week, ChatGPT Search only:

Brand	Mentions	Citations	Mention SOV	Citation SOV
You	8	3	8 ÷ 28 = 29%	3 ÷ 19 = 16%
Competitor A	11	9	39%	47%
Competitor B	6	5	21%	26%
Competitor C	3	2	11%	11%
Tracked total	28	19	100%	100%

You can be “in the conversation” on mentions and still lose the source war. That table is why the columns stay split. Repeat it per engine. If Perplexity flipped the citation column and Overviews did not, the blended average would lie about where to work.

Same 20 prompts, same week, still a teaching sheet — not a client result. Three engines, mention SOV only, so the blend is visible:

Engine	You	A	B	C	Your mention SOV	What the blend hides
ChatGPT Search	8	11	6	3	8 ÷ 28 = 29%	You are in the room, losing first-source
Perplexity	5	8	7	4	5 ÷ 24 = 21%	Closer fight; citations may be the real story
AI Overviews	2	6	4	1	2 ÷ 13 = 15%	The hole, if Google is where deals start
Blended (bad)	15	25	17	8	15 ÷ 65 = 23%	Overviews looks “fine” because chat inflated you

Citation SOV on the same invented week: ChatGPT 3 ÷ 19 = 16%; Perplexity 4 ÷ 20 = 20%; Overviews 1 ÷ 10 = 10%. A 23% blended mention number would have you celebrating chat while Google still names Competitor A. Fund the hole, not the average.

Event count versus run count — the other silent lie. One answer names You, A, and B:

Counting method	What you increment	Your mention SOV if that is the only run	Honest?
Run count	One “visible” for you	100%	No — two rivals were in the same answer
Event count (use this)	You +1, A +1, B +1	1 ÷ 3 = 33%	Yes — co-presence is the point of SOV
Winner-take-all	Only the first name	100% or 0%	No — first-name is a different KPI

Log first-name as first_mention if you care about primacy. Do not let it steal the SOV denominator.

What formula belongs on the slide?

Write the formula in words a CFO can repeat. Then say which product computed it, if any.

Formula	Math	Use when	Lie if you
DIY mention SOV	Your mention events ÷ mention events of you + locked competitors	Default for a panel you ran	Count runs instead of mention events (a co-named answer is not “one”)
DIY citation SOV	Your citation events ÷ citation events of tracked domains	Source-credit competition	Count a competitor roundup that names you as your citation
DIY recommended share	Runs where you are picked ÷ money-bucket runs	Pipeline-shaped prompts	Mix how-to prompts into the same %
Semrush Brand Performance SoV	Mentions relative to competitors; their UI also uses how high the brand appears	You subscribed and can still show raw prompts	Treat their pie as a census of the web
Semrush Enterprise AIO SoV	Mentions + position; ChatGPT also factors topic search volume (their guide)	Enterprise seats, labeled as weighted	Compare it to last quarter’s DIY count SOV without a footnote

Semrush’s simple published formula is:

(your AI mentions ÷ total AI mentions across all brands in your category) × 100

Their own example is 10 mentions out of 100 category mentions = 10%. That is an arithmetic illustration in a vendor blog, not a benchmark for your niche. Do not put “aim for 10%” on a roadmap because a table needed a number.

If a vendor score and the panel disagree, the panel wins until you can see the vendor’s raw prompts. Tools amplify a ritual. They do not replace the denominator.

Worked DIY math, still hypothetical. Twenty comparison prompts, Perplexity, you plus two locked rivals:

Row	You mentioned	A mentioned	B mentioned	You cited	A cited	B cited
Events	9	14	7	6	11	5
DIY SOV	9 ÷ 30 = 30%	47%	23%	6 ÷ 22 = 27%	50%	23%

Now a position-weighted vendor score on the same answers might move you from 30% to 22% because you were named last in long lists. That is not a traffic crash. That is a formula. If Q2 used DIY counts and Q3 used Enterprise AIO, the quarter-over-quarter “drop” is not evidence until you recompute one week on both formulas.

Do	Do not
Put the formula sentence under the percentage	Orphan a % with no prompt count
Keep DIY and vendor-weighted in separate columns	Chart them as one series
Recompute a shared week when you change tools	Declare a win because the new pie is friendlier
Show recommended-share beside SOV on money prompts	Replace SOV with “they kind of liked us”

I will not publish a fabricated studio SOV, a “typical B2B AI SOV,” or a target band. If a later case study has a dated panel, publish that panel. Until then, the method is the receipt.

What can Search Console not tell you about SOV?

Google launched dedicated Search generative AI performance reports on June 3, 2026. The Search report covers AI Overviews and AI Mode impressions. As of August 2026, that report is impression-led. It is not clicks, not queries, not ChatGPT, and not a competitor set. If someone pastes a GSC sparkline into a slide titled “AI share of voice,” stop the meeting.

Question	GSC gen-AI report	Frozen prompt panel
Were our links shown in Google AI features?	Yes — impressions	Only if you logged the Overview
Who else was named or cited?	No	Yes, if you scored it
Did ChatGPT Search cite us?	No	Yes
Which query triggered the Overview?	No query dimension	You already know the prompt
Mention without a link?	No	Mention column
Share of voice vs rivals?	Impossible — no denominator	The whole point of the panel
Clicks from the Overview?	Not in this dedicated report	Analytics, with caveats

Google’s impressions help page still matters: an Overview link counts as an impression only if it is scrolled or expanded into view, and every link inside the Overview inherits one SERP position. That is inclusion physics for you. It does not tell you Competitor A’s slice.

If the property does not show the generative-AI report yet, Google’s help page listed two reasons during the 2026 rollout: not enough gen-AI impressions, or the property was not in the set. Do not invent a workaround SOV from the classic Web performance report. Use the panel.

• GSC gen-AI impressions trended beside the panel, never instead of it
• Chart titled “Google AI-feature impressions,” never “SOV”
• Search generative AI control still set to Include unless you meant to vanish (settings help)
• Discover gen-AI is a separate report if Discover is even a channel

How to sit GSC next to the panel without renaming it SOV:

1. Export the generative-AI impression report for the same week as the panel cycle (Pages + dates).
2. Join on URL, not on hope. A cited Overview URL in the log should appear in the export if the link was scrolled or expanded into view.
3. If the log says cited and GSC is silent, check collapse/scroll rules and canonical assignment before accusing Google of lying.
4. If GSC impressions rose and Overview citation SOV on the frozen set did not, you may be getting shown on queries you are not measuring — expand the panel, do not celebrate SOV.
5. Never divide your GSC impressions by a competitor’s estimated impressions. You do not have their property. That fraction is fan fiction.

Join result	Read it as	Do not read it as
Panel cited + GSC impression on that URL	Google showed a link people could see	You won category SOV
Panel cited + GSC flat	Possible collapsed module, or too little data	Automatic tracking bug
Panel absent + GSC up	Other queries, other URLs	The frozen set is fine
Panel mentioned only + GSC silent	Expected — unlinked names are not impressions	“GSC is broken”

GSC is a Google-only inclusion lens. SOV is a competitive ratio. Keep the nouns honest.

What usually breaks the number first?

The first failure is almost never “we need a more expensive dashboard.” It is a denominator you will not defend.

Failure	What it looks like	What it costs	What you do instead
Blended engine SOV	One pie, three products	You fund the engine that already likes you	Three rows, always
Moving competitor set	A weak rival appears in week 3	Your % inflates; nothing changed	Lock 3–6 names for the quarter
Brand prompts in category SOV	“Who is [you]?” lifts the chart	Executives think you are winning recommendation	Split the series
Mention = citation	One “visible?” checkbox	Hollow name-drops look like source wins	Two columns
GSC labeled SOV	Impression line in the board deck	You optimize Google-only, miss chat	Rename the chart
One-run brief	A Thursday screenshot	Noise treated as strategy	Three cycles before a decision
Vendor black box	A 0–100 “AI Visibility” score with no prompts	Tickets nobody can audit	Panel wins until raw prompts appear
Stolen industry target	“Average SOV is X%” with no primary study on your panel	Fake urgency or fake comfort	Trend your counts

Failure mode I see in audits: the team ships ten blog posts because “SOV is low,” while the panel shows they are named constantly and almost never cited. That is not a volume problem. That is an extractable-page and corroboration problem. The wrong SOV definition bought the wrong quarter.

I will cite receipts I can defend: SEO certified since 2021, 500+ automations built, 20,000+ hours on agentic systems, 35,000+ hours saved for clients, hundreds of production sites. I will not attach a fake SOV recovery curve to any of those.

What does a week of measurement look like?

If you only have a week, you are not “doing AEO.” You are standing up a scoreboard that can survive week two. Skip the content calendar.

Day	Action	Done looks like
1	Freeze 25 prompts and 3–6 competitors	IDs, buckets, owners, locked rival list
2	Write the codebook; build the sheet	Mention / citation / recommended / accuracy columns
3	Run ChatGPT Search on the money bucket	Dated rows, Sources noted
4	Run Perplexity on the same IDs	Same
5	Manual Overview log on the same IDs (desktop, then spot-check mobile)	Cited domains, you named/cited/absent
6	Compute mention SOV and citation SOV per engine	Counts on the slide, not a lonely %
7	Three tickets from the worst bucket	URL + owner, or an explicit “not this month”

What to skip in that week:

• A new blog calendar
• llms.txt as a measurement proxy
• AI-only schema that does not match visible text
• Buying a second dashboard before the first panel exists
• A target SOV copied from a vendor index industry table

Semrush’s Index reported that in their study, the three most visible brands accounted for 82.9% of category visibility in News and Media and 41.4% in Finance. Those figures describe concentration in Semrush’s 126-million-prompt U.S. mix, January–April 2026. They are not your target. They are evidence that “good SOV” is not a portable number. If you put 82.9% or 41.4% on an internal slide, put the study name, window, and “not our panel” on the same line.

After week 1, the cadence that keeps the ratio honest:

Cadence	Owner	Artifact
Weekly	SEO or visibility lead	Money-bucket runs on in-scope engines
Weekly	Same	Two SOV columns per engine + ticket list
Monthly	Marketing lead	One-pager: counts, engines, competitor list, three decisions
Quarterly	Leadership	Re-version prompts; re-lock competitors; do not stealth-edit IDs

Miss the weekly log and you will argue from vibes again by week six. The ritual is the product.

How to read four weekly cycles without turning noise into a strategy. Same IDs, same rivals, same engines. Direction beats a single point.

Pattern across four weeks	Likely read	Ticket shape
Mention SOV up, citation SOV flat	You are getting named from third parties	Extractable owned page + corroboration
Citation SOV up, mention SOV flat	You are a source, not a brand in the sentence	Lead + entity on the cited URL
ChatGPT up, Overviews flat	Chat retrieval moved; Google did not	Do not copy the chat win onto the Overview page job
All engines down together	Competitor set or prompt demand shifted — or you stealth-edited IDs	Diff the panel version before rewriting the site
Accuracy fails on brand prompts while category SOV holds	Hallucinated offer/price/city	About / FAQ / schema agreement, not more blog posts
One-week spike, then revert	Variance on 20 prompts	Do not brief leadership off week 3 alone

• Week-over-week change shown as counts, then as SOV
• Engine columns never summed into a “company SOV”
• Panel version printed on every chart
• No target band copied from Semrush industry concentration tables

Four points is still a small sample. Treat it as a decision aid, not a census of the market.

How do you report SOV without a fake industry benchmark?

Bring one page. Not a TED talk. Not “our AI SOV is below average.” There is no average that transfers.

Slide structure

1. Object: mention SOV and citation SOV on panel v1, dates, engines, locked competitors.
2. Counts first: named in n of N money prompts, per engine, with the rival counts beside yours.
3. Ratios second, labeled DIY or vendor-weighted.
4. Hole: the engine or bucket that is actually losing.
5. Three tickets, not twenty outlines.
6. Ask: keep running the panel, or fund a visibility audit if nobody owns the log.

Exec question	Bad answer	Better answer
“What is our AI share of voice?”	“18%” with no footnote	“8 of 20 ChatGPT recommendation prompts named us; A named in 11. Mention SOV 29% on that engine, panel v1, week of [date].”
“Are we above average?”	A vendor industry table	“No portable average. Here is our three-week trend on the same IDs.”
“Should we chase ChatGPT or Google?”	The louder LinkedIn screenshot	Deal-influence × current hole on the panel
“Did the blog rewrite work?”	Sessions	Same prompts, mention vs citation, two cycles later
“Can we use Semrush’s pie?”	Yes, silently	Yes, and the prompt list it was built from, or no

Avoid asking for sympathy that “AI stole our SOV.” Ask for a denominator and an owner.

If you want a vendor number in the appendix, quote the product’s formula. Semrush Brand Performance is mention-share plus position in their UI. Enterprise AIO adds ChatGPT topic volume. Those are not the same number as DIY event-count SOV. Mixing them across quarters is how a “decline” appears that is really a formula change.

When is this not worth computing yet?

Skip the SOV program when the inputs are fiction. A ratio on garbage is worse than no ratio, because it funds tickets.

Skip SOV when	Do this first	Then start SOV
The site is not indexed or not snippet-eligible	URL Inspection, noindex / nosnippet accidents	After five money URLs fetch clean
You cannot name 3 competitors sales accepts	A one-hour sales interview	Lock the list
You are mid-rebrand or mid-merger	Freeze the legal name and About facts	After entity pages agree
Nobody will run the panel twice	Assign an owner or do not start	Week 1 and week 3 on the calendar
You only care about one branded query	Accuracy log, not competitive SOV	Keep brand prompts in their own series
Leadership wants a guaranteed Overview share	Read Google: indexing and serving are never a promise (AI features)	Measure inclusion; do not sell a slot SLA

DIY the recipe above if one named person will own the sheet. Hire when the argument is political, the engines disagree and nobody can reconcile them, or you need an outside prioritization hammer. The lane page is /visibility. The system map is the AEO playbook.

A visibility audit is the right next step when you cannot produce mention vs citation vs engine split after two weeks of honest logging — not when you dislike the first percentage.

FAQ

What is share of voice in AI answers, and how do I measure it?

It is your mention and citation slice versus a frozen competitor set on a frozen prompt panel, computed per engine. Freeze 25–40 buyer prompts and 3–6 rivals, score every run as named, cited, both, or neither, then divide your events by the tracked-brand totals. Do not import classic SEO SOV, and do not treat Search Console impressions as the competitive ratio.

How do I know if AI share of voice measurement is working?

It is working when the same prompt IDs produce dated mention and citation columns per engine, the competitor list did not move, and every material miss became a ticket. A working program shows counts a skeptic can re-run. A vendor pie you cannot map to prompts is not working, even if the percentage looks tidy.

What usually fails first when teams try this?

The denominator. Teams blend engines, fold “who is [us]?” into category SOV, or add a weak rival when the number looks bad. The next failure is collapsing mention and citation into one “visible?” checkbox, which hides hollow name-drops. Fix the codebook before you buy another dashboard.

How long does this take to show results?

A usable baseline exists after one complete panel cycle — often a week if someone actually runs it. A trend worth a budget argument needs at least three cycles on the same IDs, usually three to six weeks. Citations can lag page changes; Google does not promise indexing or serving. Do not invent a day-count SLA for Overview share.

What should I skip if I only have a week?

Skip a new editorial calendar, llms.txt theater, AI-only markup, and a target SOV copied from someone else’s industry table. Spend the week freezing prompts and competitors, running three engines, splitting mention vs citation, and writing three tickets. Measurement first. Publishing volume later.

When is this not worth doing yet?

When the site is not snippet-eligible, you cannot lock competitors, you are mid-rebrand, or nobody will run the panel twice. Competitive SOV also waits if you only care about branded accuracy — that is a different series. Stand up eligibility and an owner, then compute the ratio.

CTA

If you cannot show mention SOV and citation SOV per engine on a frozen panel, you do not have share of voice — you have a screenshot.

Lane: /visibility · Book a visibility audit.