Salary benchmarking has never had more data available — and compensation bands have never been more frequently wrong. The data sources have proliferated (Levels.fyi, Glassdoor, LinkedIn Salary, Radford, Mercer, Comp Analyst, H-1B disclosure data, state pay transparency postings), but more sources does not mean better decisions if you do not understand what each one is measuring, who submitted the data, and how to adjust it for your specific role and location.
This guide treats benchmarking as a methodology problem, not a tool problem. The tools are inputs. The methodology is what produces compensation bands that actually reflect the market you are hiring in.
Why Most Salary Benchmarking Is Inaccurate
Four structural problems that cause most benchmarking to produce unreliable results:
1. Mixing populations that should not be mixed. Levels.fyi data is heavily skewed toward FAANG and FAANG-adjacent companies. Using it to benchmark salaries at a 50-person startup produces bands that are 30-50% above the actual competitive market for a company of that size and stage. The data is not wrong — the comparison is wrong. Different talent pools require different data sources.
2. Using cost-of-living proxies for labor market data. The most common adjustment error: using cost-of-living indices (which measure what people spend) instead of labor market rates (which measure what employers pay). Austin, TX costs 40% less than San Francisco to live in. But software engineering salaries in Austin are 80-90% of San Francisco levels, because there are enough employers competing for that talent to bid wages up independent of the cost of living.
3. Stale data presented as current. Many compensation consultants and HR platforms are still selling survey data collected in 2022-2023. Tech salaries experienced significant corrections in 2023-2024 after the 2021-2022 inflation. Benchmarks from two years ago are systematically overstated for most IC roles and understated for AI/ML specializations.
4. Ignoring total compensation in favor of base salary. A $160K base at a company with no equity and minimal 401k match is not a better offer than $145K base at a company with $30K in annual RSU vesting and a 6% 401k match. The total compensation comparison is the right frame; base salary comparisons produce misleading bands.
Data Sources Ranked by Reliability
| Source | Best For | Limitations | Cost |
|---|---|---|---|
| **Levels.fyi** | Public tech company ICs (L3-L7 equivalents), verified total comp | Heavy FAANG bias; limited for non-tech companies, small companies, or senior leadership | Free |
| **Radford (Aon)** | Industry-wide comp surveys, executive and manager levels | $15,000-$30,000+/year for full access; 6-month survey cycle creates lag | Paid |
| **Mercer (TRS)** | Broad industry benchmarking, global coverage | Similar cost to Radford; best for multi-country benchmarking | Paid |
| **Comp Analyst (Salary.com)** | Mid-market companies, accessible at $3,000-$8,000/year | Less granular than Radford; methodology less transparent | Paid |
| **LinkedIn Salary Insights** | General directional data; useful for non-tech roles | Unverified self-reported; mix of total comp and base; less reliable for senior levels | Free (partial) |
| **Glassdoor** | Triangulation, candidate-facing expectations | High variance; methodologically weak for senior or specialized roles | Free |
| **H-1B LCA Data** | Prevailing wage floors for visa-sponsored roles; exact job code | Regulatory minimum, not market rate; PERM data is typically 12-18 months stale | Free |
| **Pay transparency postings** | Companies in CA/CO/NY/WA must post salary ranges; real market data | Only covers posted roles; ranges are wide; does not show where actual hires land | Free |
For pre-built tech salary benchmarks by role and level, see tech salary benchmarks by role and level — covers SDE, DevOps, Data Engineer, PM, and leadership roles across seniority levels.
How to Normalize Data Across Sources
Using multiple sources requires normalization to make them comparable. Three normalization adjustments are almost always necessary:
Normalize to total compensation, not base salary. Add employer-side 401k match, equity (using current market value for RSUs, fair value for unvested options), and target bonus to create total comp comparisons. This is especially important when comparing data from Levels.fyi (which reports total comp) against Glassdoor (which often captures only base salary).
Normalize for company tier. Levels.fyi skews Tier 1 (FAANG, top 20 public tech). Most companies compete in Tier 2 (scaling public companies, late-stage startups) or Tier 3 (early startups, non-tech companies). Apply a tier discount: roughly 15-25% below Levels.fyi benchmarks for Tier 2, 25-40% for Tier 3. The discount reflects the talent pool you are actually competing in.
Normalize for location using labor market rates. For remote work salary adjustments by city and geography, see remote work salary adjustments — it covers the cost-of-labor vs cost-of-living distinction in detail, including tier structures used by Google, GitLab, and Stripe.
Normalize for data recency. Discount survey data by approximate market movement since collection. Tech salaries declined roughly 8-12% from peak (2021-2022) in the 2023-2024 correction for most IC roles. If you are using survey data from 2022, apply a downward adjustment for most engineering roles and no adjustment or a small upward adjustment for AI/ML specializations.
Building Compensation Bands from Benchmark Data
A compensation band for a given level defines the minimum, midpoint, and maximum salaries for that role level. The standard methodology:
1. Establish the market midpoint. This is typically P50 (median) from your benchmarking data, normalized as described above. P50 means 50% of the market pays below this number for equivalent work.
2. Set band width. Standard band width in tech is ±15-20% around the midpoint. A $150K midpoint produces a band of $127K-$180K (at ±20%). The range accommodates variation in experience within the level, geographic adjustment, and hire-in negotiation.
3. Determine your target positioning. P50 means you are at market. P65-P75 means you pay better than 65-75% of competitors. Most funded tech companies target P65-P75 for base salary and use equity to reach P75+ for total compensation. Companies in highly competitive talent markets (AI research, infrastructure security) often need P75+ base to be competitive.
4. Check internal equity against the band. Sort current employees in the same role level by their current salary. Anyone below the band minimum should be flagged for immediate adjustment. Anyone at or above the band maximum is either mis-leveled or due for a level change. Salary at maximum of a band without a corresponding level change creates retention risk.
For how these bands interact with negotiation mechanics at the offer stage, see the salary negotiation guide for recruiters. For how contractor costs compare against full-time bands, see contractor vs full-time cost comparison.
Maintaining Bands Over Time
A compensation band is not a static document. Three practices that keep bands accurate:
Annual refresh cycle. Update all bands every 12 months using current-year survey data. For high-velocity roles (AI/ML, platform engineering, security), consider a 6-month refresh. Flag roles where the midpoint has moved more than 10% from the prior year for accelerated review.
Offer decline tracking. When candidates decline your offers, ask why. "Compensation didn't meet expectations" combined with a specific competing offer figure is direct benchmark data from the exact talent pool you are competing in. Aggregate this across 20+ offer declines and you have a highly targeted market signal that supplements survey data.
Pay equity audit. Run an annual regression analysis with salary as the dependent variable and role level, location, years of experience, and performance rating as independent variables. Any protected class variable (gender, ethnicity) that shows statistical significance in that regression indicates a pay equity problem in your current population — independent of whether your bands are set correctly. Tools like Syndio, Trusaic, or Lattice's compensation module automate this.
How Nextmantra AI Approaches This
Salary benchmarking produces the bands; the interview process determines which candidates within those bands you can actually close. Nextmantra AI evaluates candidates against the specific requirements of each role — technical depth, domain knowledge, relevant experience — so the compensation offer is extended to candidates who have already proven qualification, not just claimed it on a resume. That distinction matters at the offer stage: you know what you are paying for. See how Nextmantra AI handles this
Frequently Asked Questions
What are the most accurate sources for tech salary benchmarking?
For individual contributor roles at public tech companies, Levels.fyi is most accurate — it uses verified total compensation data from submitted pay stubs and offer letters. For broader industry data, Radford (Aon) and Mercer provide statistically rigorous employer-submitted survey data at $15K-$30K+/year. Using two or three sources and triangulating is better than relying on any single source.
How do you adjust salary benchmarks for location?
Use a cost-of-labor adjustment, not cost-of-living. Cost-of-labor reflects what the local talent market charges — which differs significantly from cost of living in many markets. Use Levels.fyi's location-specific percentile data, Radford metro adjustments, or LinkedIn Salary Insights location filters rather than cost-of-living indices.
How often should salary bands be updated?
Every 12 months at minimum; every 6 months for high-velocity markets like AI/ML, security, and cloud infrastructure. Bands calibrated in 2021-2022 are significantly misaligned for most roles in 2026. The risk of outdated bands is overpaying and creating internal equity problems, or underpaying and losing candidates at the offer stage.
What is the correct sample size for reliable salary benchmarking?
A minimum of 50 comparable data points per role-level-location combination. Under 50, outliers (particularly FAANG outliers on Levels.fyi) will skew the median significantly. When sample size is below 50, weight the data accordingly and triangulate with a second source.
What is the P50/P75 rule for setting offer targets?
P50 is the market midpoint — 50% of comparable employers pay below this. P75 means better than 75% of the market. Most funded tech companies target P50-P65 for base and use equity to reach P65-P75 total comp. Companies prioritizing talent density target P75+ base. Band width is typically ±15-20% around the midpoint.
How do you benchmark salaries for roles with no direct market comparison?
Decompose the role into underlying skill and experience components and weight them by role composition. A 'Platform Engineer' combining SRE and DevOps skills can be benchmarked as a weighted average of both market rates proportional to actual role composition. Document the methodology — it makes the resulting band easier to defend in offer negotiations.
How do you ensure pay equity within compensation bands?
Run a regression analysis with salary as the dependent variable and legitimate compensable factors (level, years of experience, location, performance rating) as independent variables. Statistical significance of protected class variables indicates a pay equity problem. Tools like Syndio, Trusaic, or Lattice's compensation module automate this analysis.
Conclusion
Reliable salary benchmarking requires knowing which data sources are appropriate for your company type and size, normalizing for total compensation and location using labor market rates rather than cost-of-living proxies, and building bands that are maintained annually rather than treated as permanent fixtures. The methodology is the differentiator — the tools are widely available. Getting the methodology right means fewer lost offers to compensation mismatches and fewer internal equity problems created by inconsistent banding.
Want to ensure your compensation offers reach qualified candidates faster? [See Nextmantra AI in practice](https://nextmantra.ai/platform)
Sources: Levels.fyi Total Compensation Data 2025; Aon Radford Technology Survey 2025; LinkedIn Salary Insights 2025; Glassdoor Compensation Data 2025; SHRM Compensation Survey 2025; WorldatWork Total Rewards Trends 2025; PayScale Compensation Best Practices Report 2025.