Semantic Negative Keyword Harvesting: Using Vector Embeddings to Filter Irrelevant Search Intent

The Structural Limits of Traditional Negative Keyword Scripts

For over a decade, negative keyword management has been dominated by exact-string matching scripts and regex filters. These agents scan the search term report, flag terms that literally contain a blocked substring, and append them to shared negative lists. The model is deterministic, auditable, and fundamentally brittle: it cannot recognize 'handyman license lookup' as irrelevant to a handyman services account unless the word 'license' or 'lookup' is literally present in the blocklist. Set the pattern too broad and you suffocate in-market traffic; set it too narrow and waste persists.

What exact-string and regex matching structurally cannot catch

• Misspellings and edit-distance variants: 'plumming service' versus 'plumbing service' requires a fuzzy matcher; one-character mutations evade substring checks entirely.
• Zero lexical overlap synonyms: a search for 'unclog bathtub drain' shares no tokens with the blocked term 'plumber,' yet the intent maps cleanly toward a plumbing offer — or away from it — with no string-level signal.
• Context-dependent homographs: 'apple' is a fruit query in a produce campaign and a device query in a repair campaign; string matching cannot read context.
• Emerging intent drift: seasonal terms, trend slang, and regional phrasing outpace manually curated lists built weeks earlier.
• Cross-category conflations: 'hire a plumber for a photoshoot' contains the token 'plumber' but signals a creative-services need, wasting budget when matched while being correctly classified as negative only by semantic context.

The practical consequence is a negative list that either bloats to thousands of redundant entries or misses entire intent clusters. Most accounts under $50,000/month in spend simply do not have the analyst hours to review the search term report daily, which is why manual review cycles drift to weekly or monthly, and why automated exact-match scripts create a false sense of safety. The gap is not the volume of negatives; it is the precision of intent classification.

How Vector Embeddings Represent Search Intent

A vector embedding maps a string of text into a high-dimensional numeric space, typically 768 to 3,072 dimensions depending on the model. The critical property is that cosine distance between two embeddings correlates with semantic similarity. 'cheap plumber' and 'affordable plumber' sit close together even though they share only the word 'plumber.' On the other hand, 'plumber jobs' sits far away from 'plumber near me' along the dimension that separates employment queries from service queries. That geometry is what makes vector embeddings useful for negative keyword harvesting: instead of asking whether a query contains a blocked token, you ask whether its position in semantic space is closer to your converting intent centroid or to a known irrelevant intent cluster.

Vector proximity versus string matching

Embeddings alone fix the lexical blind spots but not the reasoning gaps. A query can be semantically close to a negative cluster in embedding space while still representing a convertible micro-moment. That is why the pipeline needs a secondary LLM scoring pass: the embedding layer proposes distance-ranked candidates, and the LLM layer reasons about the match between the query's intent and the specific service description, price band, and fulfillment capability of the advertiser. This two-layer architecture is the core of modern semantic negative keyword harvesting.

Building a Semantic Negative Harvesting Pipeline

The harvesting pipeline has six stages: ingest, embed, compare, classify, score, and stage. Ingest pulls the search term report from the Google Ads API, including impressions, clicks, cost, and conversion data with a configurable lookback window. Embed converts every search term into a vector using a model fine-tuned for advertising queries. Compare computes cosine distances to a set of reference vectors — your core offering centroid, your negative intent clusters, and your proven conversion centroid. Classify hands the highest-uncertainty candidates to Gemini 3.8 Flash for natural-language reasoning. Score fuses embedding distance, LLM confidence, and historical conversion evidence into a single 0-to-1 negative score. Stage writes the candidates to the PPC Tuner review queue with an evidence trail for human decision-making.

• Ingest: pull search terms with at least 3 impressions or $2 in spend in the trailing 14-day window to keep the candidate set statistically meaningful.
• Embed: batch-encode queries and reference documents into the same latent space; normalize vectors to unit length for stable cosine math.
• Compare: compute distance to the converting centroid and the negative centroid; flag the 10-20% of queries where the two distances fall within a 0.15 margin of each other.
• Classify: evaluate flagged candidates with Gemini 3.8 Flash, supplying the client's core offering schema, price tier, and service area.
• Score: combine model outputs into a confidence score, with conversion history as a hard override — any query with a conversion in the last 60 days is pin-coded and cannot be selected for automatic rejection.
• Stage: create the negative candidate record in the PPC Tuner web app with the matched search term, suggested match type, evidence summary, and projected savings.

Intent Vector Calibration: Mapping Client Core Offerings

The reference space is everything. A semantic negative harvester trained on generic web text will classify 'what is the average cost of a plumber' as a high-intent plumbing query — and in many accounts, it actually is a qualified lead-gen query. But for a premium residential service firm with a $150 minimum job, that query converts at less than 2% and drags the account CPA above target. The intent vector must therefore be built per client, not per industry.

Building the core offering embedding set

• Service description blocks: the exact copy from the homepage, service pages, and landing pages, weighted by conversion volume.
• Converting query centroid: the embedding average of all search terms that produced a conversion in the trailing 90 days.
• Price and qualification floor: a textual boundary condition such as 'residential only' or 'minimum job size $500' that the LLM layer uses during classification.
• Asset group criteria: for Performance Max accounts, feed the asset group headlines and descriptions into the reference set so that semantic negatives align with the same offer the AI is serving.
• Negative intent clusters: pre-seeded groups for job seekers, DIY repair tutorials, wholesale suppliers, and competitor brand terms.

Calibration is where budget tier matters. A $5,000/month account might define its converting centroid from 40 to 80 conversions across 90 days — thin but workable. A $200,000/month account can cluster converting queries into service-level centroids and calibrate separate thresholds per cluster. The richer the conversion payload, the tighter the distance thresholds you can set before the LLM layer, because the embedding statistics carry more information.

Budget Tier Impact: Waste Reduction and CPA Math

The financial case for semantic harvesting comes down to recoverable wasted spend. In a meta-analysis of accounts migrated from regex-based negative lists to LLM-assisted semantic scoring, the share of clicks with zero conversions and zero relevance fell by 15 to 31 percentage points within two full conversion-lag windows. The recovery is not linear with budget because larger accounts generate longer-tail query volume where the embedding advantage compounds.

The pacing equation that governs whether recovered budget actually converts is: recovered clicks = recovered spend / blended CPC. If semantic harvesting frees $9,000 at a $6 blended CPC, that is 1,500 additional clicks. At a landing page conversion rate of 5%, that is 75 additional conversions — but only if the freed budget is redeployed into in-market keywords rather than left as unused daily budget. Put the savings on a 72-hour pacing watch to confirm delivery elasticity before locking in a higher budget ceiling.

Human-in-the-Loop: Staged Mutate Operations in PPC Tuner

PPC Tuner runs Gemini 3.8 Flash as the intent scorer and stages every proposed negative keyword mutation inside the web application for explicit human approval. The model never receives write credentials to the Google Ads account, and no mutate operation is executed until an account manager clicks approve — individually, in bulk within a confidence band, or after editing the match type. The web app maintains a complete audit trail per candidate: search term, embedding distance, LLM reasoning summary, projected wasted spend, and the approval timestamp.

• Review queue: candidates are grouped by confidence band with a per-candidate evidence panel showing the reasoning excerpt and the reference documents used.
• Bulk approve with guardrails: high-confidence negatives in the 0.85-1.00 band can be approved in bulk, but the app still pin-codes any term with a conversion in the last 60 days for individual review.
• Match-type recommendation: the LLM suggests exact versus phrase placement based on the breadth of the intent cluster, reducing the historical habit of over-broad phrase negatives.
• Revert workflow: any approved negative can be removed and the affected campaigns re-observed over the conversion lag window.
• No chat integration: all review and approval actions happen inside the PPC Tuner secure web application workspace, never through a third-party chat, messaging, or channel-based automation surface.

This staged design is the material difference between PPC Tuner and tools that push candidates directly to live accounts. Optmyzr and Adalysis offer automated negative suggestions, but their rule engines still operate primarily on string and match-type signals. Opteo focuses on automated account changes with approval modes limited to scheduled application. Ryze AI and WordStream provide advisory reports rather than staged semantic negative builds. Adzooma bundles auditing with automation but does not run intent-vector scoring against a per-client offering centroid. PPC Tuner combines the semantic sensitivity of a large language model with a human approval loop that is native to the platform and executed entirely inside the web workspace.

Operational Workflow for Ongoing Semantic Burn-Down

Semantic negative harvesting is a recurring process, not a one-time migration. The recommended cadence is daily ingestion, weekly review, and monthly list hygiene. New search terms arrive continuously, and the semantic harvester re-embeds any term that accumulates 3 impressions or $2 in spend since the last run. This keeps the candidate pool current without re-processing the entire search term history on every cycle.

Conversion lag windows and the timing of negative decisions

The most common reason a semantic negative backfires is impatience. For a lead-gen account with a 7-14 day median conversion lag, a search term that generated clicks four days ago may already be mid-funnel. The scoring pipeline must therefore separate 'no conversions yet' from 'no conversion evidence in the observation window.' PPC Tuner refuses to stage a negative for any term that shows a conversion, assisted conversion, or view-through conversion within the trailing 60 days, and the LLM classification prompt explicitly ignores the last 7 days of conversion data for click-date terms younger than the median lag.

• Day 1: export the prior month's search terms; reconcile the negative list against the converting query centroid; archive removed negatives.
• Day 2-4: run the weekly review queue; approve high-confidence bulk batches; route mid-confidence candidates to the accountable account manager.
• Day 5: push approved negatives to the account via the Google Ads API; log the mutation with the evidence IDs.
• Day 6-14: observe delivery and impression share on the affected campaigns to confirm no search-term bleed.
• Day 15: run the Lost Impression Share Calculator to verify the negatives did not reduce eligible auction coverage.
• End of month: measure wasted spend before versus after; compile the CPA impact report using the budget tier table above.

Measuring Impact, Quality Metrics, and Iteration

The success of a semantic negative harvesting program is measured with five numbers: wasted spend rate, true negative rate, false negative rate, impression share recovery, and account CPA. Wasted spend rate is the share of clicks that produced zero conversions or zero engaged sessions where the search term's semantic distance from the converting centroid exceeds the 0.65 threshold. True negative rate is the share of staged candidates that a human approves and that subsequently produce zero conversions over two full lag windows. False negative rate is the share of approved candidates that produce a conversion anyway — the number that must stay below 1% for the program to remain safe.