Lookalike segments in Demand Gen 2026: seed lists, 2.5 / 5 / 10% reach tiers and when they beat optimized targeting
Demand Gen is good at spending budget and bad at telling you who it spent it on. Optimized targeting finds an audience by itself, but it also makes the decision for you — and it will happily drift toward the cheapest, easiest-converting pocket of people who were already going to buy. Lookalike segments in Demand Gen are how you take that decision back: you define who the audience should resemble, and how far the system may wander from that definition.
Here is the working mechanic for 2026: seed list requirements, what the 2.5 / 5 / 10% reach tiers actually mean in people, what changed when the tiers became suggestions rather than constraints, how to build a seed that produces incremental results, and how to prove the lookalike beat automation instead of just reshuffling impressions.
Lookalike segments in Demand Gen vs optimized targeting
A Lookalike segment is built from a “seed” — your own first-party list — and targets people with similar characteristics. The difference from optimized targeting is where the decision originates: a Lookalike works from your definition of a valuable user, while optimized targeting optimizes against the campaign’s conversion signal and decides on its own who to add. The second mechanism is covered in detail in optimized targeting and audience signals.
The practical consequence is that the two fail differently. Optimized targeting fails toward people who are too familiar — buyers you would have captured anyway. A Lookalike fails toward the quality of your seed: a junk seed produces a junk audience no matter how much the system optimizes it.
Seed list requirements
| Parameter | Value |
|---|---|
| Number of seed lists | up to 10 at once, at least one required |
| Allowed sources | website visitors, app users, customer lists, Google Analytics audiences, YouTube audiences |
| Minimum size | more than 100 active matched people across all seed lists combined (a separate Help page states the old hard 100-user minimum no longer applies — check what your account UI shows) |
| Refresh cadence | every 1–2 days |
| If the requirement stops being met | the segment stops targeting new users; takes up to 3 days to take effect |
| Campaign type | Demand Gen (Video campaigns have their own documentation and rules) |
“Active matched people” is not the size of your database. Only part of an uploaded customer list matches, and only the matched portion counts toward the threshold. Membership and activity thresholds are covered in Audience Manager and the 100-user threshold; list preparation and upload mechanics are in Customer Match and first-party data.
Reach tiers: 2.5 / 5 / 10% — translated into people
- Narrow — roughly 2.5% of users in your target location, most similar to the seed.
- Balanced — roughly 5%.
- Broad — roughly 10%, largest reach with the least similarity.
Convert the percentage into people before you launch; it answers most budget questions in advance. If the system sees, say, 8 million active users in your geography, Narrow is on the order of 200,000 people and Broad around 800,000. On a daily budget that buys a few thousand impressions, a Narrow segment burns through frequency within days — and the rising CPM you then see is audience exhaustion, not audience quality. Controls for that are in frequency capping and target frequency.
Directional rule: the smaller the budget, the broader the similarity tier should be. Narrow only pays off when the budget can sustain creative rotation against a small pool without frequency burnout.
The 2026 change: the slider became a suggestion
Historically the reach tiers acted as a hard similarity cap. Starting March 2026, Google has been phasing Demand Gen Lookalike segments into suggestion mode: the selected tier and the seed list are used as signals, and the system may serve to qualified users beyond the 2.5 / 5 / 10% thresholds. What follows from that:
- Reach is no longer predictable from the formula. When planning frequency and budget, assume the real pool can be wider than the stated percentage.
- Reporting flags it. Results appear in the Lookalike segment row marked with a “signal” tag — that tag is how you know you are in suggestion mode rather than classic constraint mode.
- You can opt out. Opt-out runs through a form and removal takes about 3 days; UI controls arrive later in the year. If your buying model depends on a strict similarity ceiling, opting out is reasonable.
- Demand Gen only. Lookalike audiences in Video brand campaigns and in Display & Video 360 are unaffected.
- Overlap with optimized targeting. With both enabled, delivery shifts toward the Lookalike segment, though optimized targeting may still reach users beyond it.
One more detail worth knowing: with Lookalike segments, delivery concentrates on the Lookalike audience whether audience expansion is on or off — so enabling expansion “just in case” accomplishes nothing. For Video brand campaigns, the documentation recommends disabling it outright.
Building a seed that actually works
Seed quality beats seed size. Five configurations that tend to produce real lift:
- Purchasers from the last 30–90 days. The most direct intent signal, and a short window keeps the seed from filling with stale profiles.
- Top 20% by lifetime value. If you have an LTV calculation, seeding from the top quintile produces a materially different audience than “all buyers” — you are looking for people like your profitable customers, not like any customer.
- Repeat purchasers. For subscriptions and consumables this is a far better proxy for retention than a single transaction.
- Engaged YouTube viewers. Useful when transactions are too sparse to meet the size threshold on their own.
- A high-intent GA4 segment — reached checkout and did not buy, for instance. How those are built is covered in GA4 predictive audiences.
What not to do: seed from all website visitors over 540 days. That list describes your traffic, not your customer, and the Lookalike will faithfully find more passers-by. Equally, avoid mixing fundamentally different roles into one seed — a B2B procurement contact and a retail shopper average into a profile that resembles neither. If what you actually need is behavioural or search-intent targeting, custom segments are the better tool.
CRM seeds: data prep and where the volume disappears
A customer list is the strongest kind of seed and the one most exposed to technical loss. Far more is lost between the CRM export and the “active matched users” number than most teams expect. Where it goes:
- Normalisation. Stray whitespace, mixed casing, phone numbers in several formats — all of it depresses match rate. Normalise before upload rather than hoping the system copes.
- Contact quality. Role-based addresses, disposable mailboxes and mistyped numbers simply do not match.
- Recency. Three-year-old contacts match noticeably worse than fresh ones; people change email addresses and phone numbers.
- Incomplete records. The more identifiers you supply per person — email, phone, name, country — the higher the match probability. Uploading email alone is a voluntary loss of reach.
- Refresh discipline. A one-off upload decays quickly. Schedule recurring uploads or the seed starts ageing on day one.
The planning implication: do not size your seed by rows in the export. Size it by the matched active users the interface reports, and build in headroom — especially for a Narrow tier, where every lost thousand records directly shrinks the addressable pool. Upload mechanics and the fields that raise match rates are covered in Customer Match and first-party data.
Privacy and consent are a separate thread. Only upload data you have a lawful basis to use, and treat that as a legal question rather than a technical one, since the rules differ by market. The practical rule: build seeds from sources where consent is recorded, and make sure your refresh process can remove records on request.
When a Lookalike beats automation — and when it doesn’t
| Situation | Choose | Why |
|---|---|---|
| Clean purchaser list of 1–2k+ | Lookalike | The seed encodes value, not just a visit |
| Few conversions, shaky tracking | Lookalike | Automation has nothing to learn from; the seed is an anchor |
| High, stable conversion volume | Optimized targeting | The signal is strong enough and it will find more |
| Launching a new market cold | Lookalike seeded from a comparable market | Transfers a profile instead of starting from zero |
| Pure reach and brand awareness | Neither, as the primary lever | Format, frequency and creative decide the outcome |
| Narrow professional product | Lookalike, Narrow tier | Similarity matters more than scale |
Setup, step by step
- Build and upload your seed lists (up to 10) and confirm the combined matched active users clear the threshold.
- Create the Lookalike segment in Audience Manager, selecting the seed and the similarity tier.
- Allow 1–2 days for the segment to populate — that is the refresh cadence.
- Run it in its own Demand Gen campaign or ad group rather than mixing it into an existing one, or you will never separate the results.
- Review audience expansion and frequency settings.
- Produce creative for a cold audience — these people do not know you. Format requirements are in Demand Gen creative for Shorts and Gmail.
- Record the launch date in your change log; you will need it when a metric moves.
If you are still moving legacy Display campaigns into this campaign type, the migration order and the usual losses are documented in migrating Display campaigns to Demand Gen.
The first two weeks: what to watch and when to decide
The most expensive mistake with Lookalikes is judging them on day three and switching off a segment that has not finished gathering data. A realistic timeline:
| Period | What is happening | What to watch | Decisions that are fair |
|---|---|---|---|
| Days 1–2 | Segment populates | Audience size, segment status | Technical only: delivery, link correctness |
| Days 3–5 | Volume ramps, bids calibrate | Impressions, frequency, CPM | Frequency adjustments if it climbs faster than planned |
| Days 6–10 | Meaningful click volume | CTR by creative, watch depth | Cutting clearly failing creatives |
| Days 11–14 | Conversions accumulate | CPA, share of new users | Widening or narrowing the similarity tier |
| After 14 days | Enough data to conclude | Lift against control | Scale or stop |
Watch the share of new users specifically. If the Lookalike mostly delivers people already sitting in your remarketing lists, you are paying to reach an audience you already own. That is the earliest sign the seed was built from material that is too warm.
A worked comparison: two seeds, one product
A hypothetical retailer runs two parallel Demand Gen campaigns, same budget, same creatives. Only the seed differs:
- Campaign A. Seed: all site visitors over 540 days, roughly 400,000 records, Broad tier. Result: plenty of cheap clicks, a low share of new users, CPA running about 1.5× the target. The audience genuinely resembles your traffic — that is, people who arrive and leave.
- Campaign B. Seed: purchasers from the last 90 days, top value quintile, roughly 6,000 records, Balanced tier. Result: more expensive clicks, a higher share of new users, CPA inside target. The audience resembles people who pay.
The transferable lesson: six thousand right records beat four hundred thousand random ones. Seed size affects the stability of the model; seed quality affects the economics. Do not take that to the extreme, though — a 150-person seed will technically clear the threshold and produce an unstable segment that visibly drifts between refreshes.
Where Lookalikes sit in the rest of your structure
A Lookalike does not replace your other audiences; it occupies the upper-middle of the funnel. The working stack:
- Remarketing — people who already engaged. Cheapest return, limited volume.
- Customer Match — existing customers for upsell and retention.
- Lookalike — expansion to similar people, governed by your definition of value.
- Optimized targeting — discovery beyond your lists, where conversion volume supports learning.
Keeping these layers in separate campaigns or ad groups is non-negotiable. Run remarketing and a Lookalike in the same ad group and the cheap conversions from a warm audience will mask the real economics of the cold one. In reporting that looks like “the Lookalike is performing brilliantly,” when in fact you are looking at remarketing under a different label.
Proving the Lookalike produced lift
Comparing “CPA in the Lookalike campaign vs CPA in the campaign next door” proves almost nothing: audiences overlap, attribution reassigns credit, and seasonality moves both. Three methods that give an answer you can defend:
- Geo holdout. Split markets into test and control and read the difference in orders, not in reported conversions.
- Conversion Lift. The correct instrument for “would these conversions have happened anyway” — see Conversion Lift for Search and Performance Max.
- Brand Lift, when the goal sits at the top of the funnel — see Brand Lift and Search Lift on YouTube.
One caveat about the reported numbers themselves: several Demand Gen metrics were redefined during 2026, so a jump in view-through conversions may have nothing to do with your audience. How to tell the two apart is covered in today’s companion piece, the metric changed, not your results.
Six mistakes that eat the result
- Seeding from all traffic. It describes visitors, not customers.
- Narrow tier on a small budget. Frequency burns out in days, CPM climbs, and the team concludes the audience is bad.
- Running the Lookalike inside an existing campaign. Results blend and no decision is possible.
- A forgotten seed. A list that stops receiving new records becomes an archive within a quarter — monitor inflow.
- Reusing remarketing creative. Assets that worked on warm audiences usually fail here; the viewer has no context.
- Judging reported CPA after three days. Segments refresh over 1–2 days and requirement changes take up to 3 days to apply — any conclusion inside a week is premature.
Pre-launch checklist
- Seed built from valuable actions rather than all traffic; matched active users clear the threshold.
- Chosen tier converted into a headcount and reconciled with budget and target frequency.
- Suggestion-mode behaviour accounted for; opt-out submitted if a strict ceiling is required.
- Lookalike isolated in its own campaign or ad group.
- Creative produced for a cold audience.
- A lift measurement plan in place — geo holdout or Conversion Lift — not just a CPA comparison.
- Launch date for the Lookalike segments in Demand Gen recorded in the change log.
More account-level breakdowns are published in the PPC Rebels blog, and the platform fundamentals are collected in the PPC Rebels Google Ads guide.
FAQ
How many people does a seed list need?
The documentation for Lookalike segments as a targeting constraint requires more than 100 active matched people across all seed lists combined. A separate Help page states the old hard minimum of 100 users no longer applies. Go by what your account UI shows and build in headroom — only part of an uploaded list matches.
What do 2.5%, 5% and 10% mean?
They are the share of users in your target location the segment aims at: Narrow targets about 2.5% most similar to the seed, Balanced 5%, Broad 10% with the lowest similarity.
Is it true the slider no longer caps reach?
In Demand Gen from 2026, the tier and seed are used as signals and the system may serve beyond the stated percentage. Classic behaviour is preserved if you opt out; removal takes about 3 days.
How many seed lists can I use?
Up to ten at once, at least one required. Sources include website visitors, app users, customer lists, Analytics audiences and YouTube audiences.
How often does a Lookalike segment refresh?
Every 1–2 days. If the seed stops meeting requirements, the segment stops adding new users, and that takes up to 3 days to take effect.
Should I enable audience expansion alongside a Lookalike?
There is no point: with Lookalike segments, delivery concentrates on the Lookalike audience regardless of the expansion setting. For Video brand campaigns the documentation recommends turning it off.
Lookalike or optimized targeting?
With sparse or unstable conversion volume and a solid first-party list, Lookalike. With high, stable conversion volume, automation usually finds more. With both enabled, delivery shifts toward the Lookalike segment.
Can I use Lookalikes in Video campaigns?
Lookalike audiences exist for Video campaigns, but they have their own documentation and rules; the 2026 suggestion-mode change applies to Demand Gen and does not affect Video brand campaigns or DV360.
How do I know the Lookalike worked instead of just absorbing impressions?
Only by measuring lift: a geo holdout or Conversion Lift. CPA comparisons between campaigns cannot answer this because of audience overlap and attribution.
What if Lookalike conversions come in expensive?
In order: check seed quality (valuable actions, not all traffic), widen the tier if the budget is small, review frequency, and refresh creative for a cold audience. Do not change all four at once, or you will never know which one helped.
