What AI visibility target should you commit to for 2027? A plan that survives model changes

Why is a 2027 AI visibility target harder to set than an SEO target?

Because the thing you measure moves even when you do nothing. An SEO target sits on rankings that change slowly and are measured the same way each month. An AI visibility target sits on answers that vary from run to run, differ by engine and change when a model is replaced. A target that ignores those three facts will be missed or met for reasons nobody controlled.

Model replacement is the clearest case. OpenAI's ChatGPT release notes record GPT-6 Sol becoming the model for paid tiers from 7 October 2026, with GPT-6 Luna reaching Free and Go users from 8 October 2026.¹ A brand that set a target in September was measured against one model and will be judged against another. Nothing in its content changed on those dates.

November is when most marketing teams write next year's numbers. This guide sets out how to write an AI visibility target that a finance director can accept, an agency can be held to, and a model release cannot quietly break.

Prior work and a disclosure

Sigzen AI sells AI visibility audits and programmes at published prices, and part of our programme fee is linked to a share-of-answer target, so we have a stake in how targets are written.¹² The method below works with any vendor or with a spreadsheet.

Others have covered parts of this. Otterly, which sells AI search monitoring, published a stability study in September 2026 comparing how much brand mentions and cited sources move from day to day.² MaxAEO, which sells AI visibility measurement, published guidance in July 2026 on sample size and why repeat runs of one prompt are clustered rather than independent.³ What this guide adds is the planning layer on top: which number to commit to, how wide its range should be, and what the contract says when the engines change.

Pick the measure before you pick the number

Commit to share of answer: the share of answers to your buyer prompts that name your company. It is the measure closest to what a buyer sees, and the one you can move with on-site and off-site work. The other candidates each fail as a headline for a specific reason, though most belong somewhere in the report.

Why not referral traffic?

Referral traffic is too small, and it misses most of the effect. Across 101,574 websites in 250 countries, SE Ranking found that AI search sent 0.32% of all visits in January to April 2026.⁴ A target on a line that small swings on a handful of visits. Worse, the visits that follow an AI answer mostly arrive by another route.

Similarweb, which sells traffic and AI visibility data, followed US desktop users in its panel from July to December 2025. Brands recommended by ChatGPT were 2.5 times more likely to be visited within seven days, and 55.9% of those visits arrived through search, with only a small share as visible AI referrals.⁵ A referral target would credit search with work the AI answer did.

Why not citations alone?

Citations are the diagnostic, not the target. They tell you which pages and sources an engine leaned on, which is how you decide what to fix. They also move more from run to run than mentions do, and engines differ widely in how many sources they show. Our note on mentions and citations as KPIs sets out why we put mention share on top and the citation map underneath.

Three lines for a 2027 plan, and which one carries the target

Input targets

• Pages rewritten to answer buyer prompts
• Third-party listings corrected
• Facts audit closed each quarter
• Fully in the team's control

Share of answer

• Fixed, versioned prompt set
• One line per engine, with a range
• Benchmarked against named rivals
• The number you commit to

Business signals

• AI referrals in analytics
• Branded search demand
• "How did you hear about us" answers
• Tracked, not targeted

The case for a target at all

Buyers already act on these answers, so leaving AI visibility without a target leaves it without an owner. In G2's survey of 1,076 B2B decision makers in North America, EMEA and APAC in March 2026, 51% said they start software research in an AI chatbot more often than in Google.⁶ In the same survey, 69% said an AI chatbot had surfaced information that led them to choose a different vendor than they expected.⁷

Those are software buyers, and your category may differ. Still, the direction is plain enough to plan on. A line in next year's plan with a number, an owner and a review date is how work gets funded. A vague goal to "improve AI presence" is how it gets cut in the second quarter.

Planning inputs for 2027

Small in clicks, large in influence

0.32%of all website visits came from AI search, Jan–Apr 2026SE Ranking, 101,574 sites ⁴

2.5xmore likely to be visited within seven days when ChatGPT recommended the brandSimilarweb, US desktop, Jul–Dec 2025 ⁵

51%of B2B decision makers start software research in an AI chatbot more often than in GoogleG2, Mar 2026 ⁶

Fix the prompt set, then version it

The prompt set is the ruler, so it has to stay the same length all year. A target set on one list of buyer questions and judged on another is not a target. Agree the prompts before the baseline is run, write down who approved them, and freeze them for the period the target covers.

Prompts still need to change. Buyers adopt new wording, you launch a product, a rival enters the market. Handle it the way a finance team handles a restated account: give the set a version number, run old and new versions side by side for one measurement cycle, and report the target on the version it was set on. New prompts get their own line until the next planning round.

What goes in the prompt set

Buyer questions in your category, written the way buyers type them, across the stages where a shortlist forms. Leave out prompts that name your brand, because they measure recall of a name the buyer already has. Our guide to building a prompt set from buyer questions covers the sourcing. For planning, the rule that matters is size: more prompts buy more precision than more repeats of the same few.³

Our view: if a vendor cannot show you the prompt list behind your target, you do not have a target. You have a number the vendor can move by editing a spreadsheet.

Set one line per engine

A single blended number hides the fact that each engine rewards different work. The engines choose and display sources on their own rules, so the same effort moves them by different amounts. Setting one line per engine lets you see which work is paying off and stops a strong engine from masking a weak one.

The differences are large. Semrush's 2026 AI Visibility Index, built on 126 million US prompts from January to April 2026, found ChatGPT cited about 15 sources per response against about 3 for Gemini.⁸ Profound, which sells AI visibility software, analysed 11.84 billion citations from 16 April to 16 July 2026. Gemini sent 69% of its citations to company-owned sites and ChatGPT 47%.⁹

Read together, those figures suggest that owned-site work moves Gemini more directly, while ChatGPT draws more on what others publish about you. A plan with one target cannot express that; a plan with five lines can.

Share of citations going to company-owned sites, by engine

Gemini69%

ChatGPT47%

Which engines get a line depends on your buyers. Sigzen AI measures five: ChatGPT, Gemini, Microsoft Copilot, Google AI Overviews and Google AI Mode.¹² Some teams commit to two or three and track the rest. That is a sound choice, provided the engines left out are reported, so a collapse on one of them is still seen.

How wide should the target range be?

Wide enough that chance alone rarely carries you across it. Share of answer is an estimate from a sample of answers, so it comes with a range. If the target sits inside the range of your baseline, a quarter can "hit" it through noise and the next quarter can "miss" it the same way. The target is only meaningful once it sits clearly above the baseline's range.

The range narrows with more prompts and more separate runs, and widens when you slice the data thinly. A per-engine line has fewer answers behind it than the total, so its range is wider. Plan for that rather than discovering it in the first quarterly review. Sigzen AI's own reports show share of answer with a 95% range, from 150 buyer prompts run on three separate days.¹²

Our view: write targets as "the range moves above X" rather than "the estimate reaches Y". It reads as a little less ambitious in a planning deck and is far easier to defend in a review.

What happens to the target when an engine changes its model?

You re-base it, by a rule you wrote before it happened. Model releases change answers on their own, and they are frequent enough that a calendar year will see several. A target that has no re-base rule turns each release into an argument about whether the team or the engine moved the number.

The October change is a good test case. Ahrefs, which sells AI visibility tools, studied 66.6 million free, logged-out ChatGPT responses on GPT-5.6 Luna in August and September 2026, across all countries. It found 14.4% triggered a web search, so most answers in that sample came from what the model already held.¹⁰ A month later, Free and Go users moved to GPT-6 Luna.¹ Answers that come from the model's memory can shift when the model is replaced, whatever you published.

A re-base rule you can write into the plan

Name the trigger: a change of default model on an engine with its own line. Measure the full prompt set on that engine in the two weeks after the change. Keep the absolute target only if the new baseline sits inside the old range; otherwise move the target by the same distance the baseline moved. Record each re-base in the report with its date and cause. Our guide to measuring before and after a model release covers the measurement itself.

From model release to a re-based target

1. Release noted
2. Full set re-run
3. Baseline compared
4. Target moved or kept
5. Change logged

Free signals to track beside the target

The engines' own reports are partial, but they cost nothing and they show trend. Bing Webmaster Tools added an AI Performance report in public preview in February 2026, covering Microsoft Copilot, AI-generated summaries in Bing and select partner integrations.¹¹ It counts how often your pages appear as sources, which makes it a citation signal rather than a share-of-answer measure.

Put these beside the target, not in place of it. They answer narrower questions: is the engine reading our pages, and which ones? Analytics adds AI referrals, and a "how did you hear about us" field on forms adds what buyers say themselves. Our guide to AI search numbers in a board report covers which of these belongs on the slide.

Our view: one engine-reported signal per engine you care about is enough. More than that and the team spends the first week of each month reconciling reports that count different things.

How should the target sit in an agency contract?

As a written clause with its measurement attached, agreed before work starts. A target that lives only in a kick-off deck is a hope. The clause should name the prompt set and its version, the engines, how often each is run, how the range is reported, and the re-base rule. Without those five items, nobody can say afterwards whether the target was met.

Fee-linked targets raise the stakes. In Sigzen AI's Growth and Scale programmes, 20% of the monthly fee is at risk against the share-of-answer target set at the audit, and is invoiced only for a quarter in which it is met.¹² That works only because the measurement and the target are written into the agreement first.

Illustrative. Clause list used by the Sigzen AI editorial team; adapt the wording with your own counsel.

A planning calendar for the next 90 days

Most of the work happens before the target is written. A baseline needs a few weeks, a rival set needs agreement from sales, and the contract needs both. Starting in November leaves time to commit to a number in January that rests on measured data rather than a guess.

From baseline to signed target

1. November

   Agree the buyer prompt set and the rival set with sales; freeze version 1.

2. Late November

   Run the baseline on every engine you will report, on separate days.

3. December

   Read the ranges per engine; choose which lines get a committed target and which are tracked.

4. January

   Write the targets as ranges, add the re-base rule, and sign them into the plan or contract.

5. Each quarter

   Report estimate, range and any re-base; review prompts and plan version 2 if needed.

Choose the rival set with sales

A share of answer means more when it sits beside the companies your buyers compare you with. Ask sales which names come up in lost deals and in the last shortlist they saw, then hold that list fixed for the year. A rival set chosen by marketing alone tends to include the market leader and miss the firm that actually wins your deals.

What to tell finance when they ask what the number buys

Tell them what it measures, what it does not, and what the business signals beside it show. Share of answer measures how often buyers asking your category's questions see your name. It does not measure revenue on its own. The business signals, AI referrals, branded search and what buyers say on forms, show whether the visibility is turning into demand.

Keep the commitment modest and the reporting honest. A target that sits clearly above its baseline range, measured on a fixed prompt set, with a written re-base rule, is a promise a team can keep. A dramatic number on a blended score, measured however the tool happened to run that month, is one it cannot.

Our view: the best 2027 AI visibility target is a boring one. It names its prompts, shows its range, splits by engine and says in advance what happens when the engines change. Everything else in the plan can then be argued on its merits.