Try Valley
Put your outbound workflow to work.
Saniya Sood
Published:
Updated:
Published September 25, 2026 · Updated October 8, 2026
The answer up front: personalization at scale is worth paying for when every message gives the right person a credible reason to talk. A first name or a compliment does not; neither does a research paragraph that never connects to the offer. This page gives you a weighted six-dimension rubric, one annotated example, and an editing-cost measure, so you can score any tool's output the same way twice and compare vendors on evidence.
Disclosure: Valley makes an outbound platform. The rubric is Valley's own working method, not a validated instrument or a completed benchmark, and the examples are from Valley's demonstration workspace or written for this page; none are messages sent to real prospects.
How we know this category: Valley runs outbound for 1,000+ LinkedIn accounts and has sent 30M+ messages.
Valley is an AI outbound platform that finds the people most likely to buy from you - from live buying signals, natural-language search, or your own lists - qualifies them against your ICP, researches each one, and sends personalized messages across email and LinkedIn that get 15-45% reply rates, against a category average of about 2%.
► Want to see it instead of read it? Try Valley free for 7 days
What personalization has to prove
A personalized message should give a specific person a reason to consider a relevant conversation. Judge the reasoning behind the message and the effort needed to approve it, not how different it looks from the previous one. Length of research, number of personalized tokens and sophistication of language are ornament; the question is whether the preparation supports a useful, credible message.
Fix the inputs first
Build a small evaluation set before any vendor demo: twenty records is a manageable start, not a representative sample. Include straightforward fits, borderline fits, stale signals, missing information and clear exclusions, and avoid a set made only of people with rich public profiles.
Give every candidate the same offer, target criteria, exclusions, approved product facts and writing examples, with the same research cutoff date and channel requirement. If one tool uses extra licensed data, record the difference instead of pretending the inputs were identical; it may be a real advantage, but it changes what you measured.
Test learned style separately from one good draft: supply approved writing examples, then inspect new messages for different prospects without restating the style in each prompt, and record whether the product keeps examples, a style document, sent-mail context or learned edits. For research, require inspectable sources behind substantive findings rather than profile fields dressed up as a dossier. Ask for the original output before editing, keep the sources and configuration, and keep the rejected records: a tool that declines to invent an opener should not lose points for returning less copy than one that confidently fills every blank.
Score six dimensions with a shared definition
Score each record 0 (unacceptable), 1 (substantial correction needed), 2 (usable with minor edits) or 3 (approve as written). Two reviewers score independently before discussing disagreements. The weights are a starting proposal; set them before reviewing outputs if your priorities differ.
Dimension | Weight | What a strong result shows |
|---|---|---|
Factual grounding | 30% | Material claims match accessible, current evidence |
Buyer relevance | 25% | The observation relates to this person's role and the offered value |
Reasoning restraint | 15% | The message does not turn a signal into an unsupported conclusion |
Voice and clarity | 15% | Wording follows the supplied style and reads naturally |
Useful next step | 10% | The request is specific, proportionate and easy to understand |
Context restraint | 5% | Professional context without needless personal detail |
Weighted score = the sum of each dimension's score divided by three, multiplied by its weight. Report the dimensions alongside the total, because an attractive average can hide an invented fact: any material fabrication, wrong identity or breach of the exclusion rules fails the record regardless of total.
One annotated example
Valley's demonstration workspace uses an example prospect, Maya Chen, VP Revenue at Fieldwork, with a hiring signal. The draft reads: "Saw Fieldwork is hiring AEs. That usually means new lists to build and first touches to write. Valley finds the prospects and drafts outreach from real research. Want to see a few examples for Fieldwork?" These names belong to the demo, not to a customer.
There are three kinds of statement in it. "Fieldwork is hiring AEs" is the factual premise and needs a matching source in a real evaluation. The second sentence is an inference about the work that may follow hiring. The last two connect that possible work to an offer and a modest request. Score each layer separately. The restraint can improve: hiring does not establish that Maya needs outside help, and an editor could soften the move to "If the new team needs help building its first lists, I can put together a few researched examples." That rewrite is an illustration for this page, not a measured gain in replies.
Compare a deliberately weak example, written for this page: "Congratulations on your incredible growth. Your team must be overwhelmed, and Valley will solve your pipeline problems." The compliment adds nothing, "overwhelmed" invents a condition, and the promise exceeds the evidence. No amount of first-name substitution repairs that.
Give your sales team
an unfair advantage.
Measure the cost of reaching an approved draft
Alongside scores, record active review time, factual corrections, source checks and substantive rewrites, and distinguish a changed comma from a replaced reason to contact someone. Useful measures: the share of records approved without substantive edits, the share with unsupported material claims, median active review time, and the share correctly excluded. Define each denominator; a system that rejects several unsuitable prospects should not be described as producing fewer useful drafts without crediting the correct exclusions. Record reviewer disagreements before reconciling them, fix the brief when the brief was the problem, and rerun every candidate under the same revised conditions.
Keep writing quality separate from sending results
A strong offline draft is a prerequisite, not proof of replies or meetings. Audience selection, offer, timing, sender reputation and follow-up all move a live campaign, so do not credit personalization for a difference in replies when several of those changed together. If you later run a live experiment, define its primary outcome and review window in advance, track meaningful positive replies separately from all responses, keep delivery failures and exclusions visible, and use a statistical method suited to the sample before making comparative claims.
Keep the evaluation reusable
Save the record identifier, source URL and date, configuration version, original draft, reviewer scores, edit time, final decision and reason. Store commercial or personal information only where your normal access controls allow. Refresh stale examples and keep a fixed subset so change stays observable over time. Clay, Amplemarket and lemlist all document research or contextual writing worth scoring alongside Valley: see Valley vs Clay, Valley vs Amplemarket and Valley vs lemlist, and use the delegation guide to inspect approval boundaries.
FAQ
What is personalization at scale? Producing a specific, credible reason to talk for every prospect in a campaign, not inserting a first name or a compliment into a template. The test is whether each message is grounded, relevant and restrained, at the volume you send.
How do I score personalized outreach from an AI tool? Use the six-dimension rubric above on twenty records with two reviewers, fail any record with a fabricated fact, and record the editing cost to reach an approved draft.
What is a good personalized cold email? One whose factual premise has a source, whose inference is stated as an inference, and whose ask is proportionate. The annotated example above shows the three layers.
Does better personalization raise reply rates? Often, but a draft test cannot prove it. Audience, offer, timing and sender reputation move replies too, so measure writing quality offline and sending results in a separate, pre-registered experiment.
How is this different from a vendor's own benchmark? It is a method you run on your own records with your own inputs. Vendor benchmarks use their records and their inputs, which is why this page publishes no reply-rate claims from it.
How this page was made: the facts were read from the sources named in the text on October 8, 2026; drafted with AI assistance; fact-checked, reviewed and edited by Saniya Sood.
