How to read a salary benchmark without being misled

A salary benchmark arrives looking like a fact and behaves like an argument. Before you let one set a band for a senior IT, AI or data seat, ask the two questions that decide whether it's telling you about your market or someone else's. First, the lag: when was this data actually collected, and how long did compilation take? Second, the basket: which roles dominate the sample, and are they anything like the one you're pricing? Neither question is answered by the number on the page. Both are answered by the method note, which is exactly why the honest surveys publish one and why it's the first thing to read.

The number itself is the last thing that should move you. A benchmark is a summary of a method, and a summary is only as trustworthy as the process behind it. Read the process, and the range becomes what it actually is: a reference point with a known set of blind spots you can correct for. Skip the process, and you've imported someone else's blind spots into your budget without knowing they're there.

Start with what a benchmark actually is. A comp house collects pay submissions from participating employers on a fixed annual cycle, cleans them, matches each one to an established job code, ages the data to a common reference date, aggregates it into percentiles, and only then publishes. Every step is deliberate, and the scale is genuine: Aon's 2025–26 India survey draws on more than 1,400 organisations across 45 industries, and Mercer's Total Remuneration Survey covers over 8,000 roles at more than 1,500 companies. That scale is what makes the output authoritative. It's also what makes it a description of the past rather than the present, because you can't aggregate 1,500 companies quickly.

So a published band answers a precise question: across many employers, what did roles matched to this code pay, as of the reference date? That's a useful question. But it isn't what the specific person you're trying to hire will sign for, and it isn't what your competitor quoted them last week. The benchmark is true on average and precise about no one, and the two things that pull a specific offer away from that average are the two things the method quietly encodes: how old the data is, and which roles filled the basket.

The lag is built into how a survey is made. Each step between submission and publication adds distance between the number on the page and the number being signed. In a steady segment that distance is harmless. In a re-pricing segment, which is most of the ones a GCC actually competes in, the lag runs in exactly the direction that hurts you, because the roles that pushed the market are the ones that closed at figures the compilation never captured.

You can see the movement in job postings before you see it in the compiled bands, because a posting is real-time and a survey is not. Naukri's JobSpeak index recorded AI/ML hiring up around 45% over the year, with the senior 20-plus-LPA band up about 16%, a leading signal by definition. Behind that sits real scarcity: the Deloitte–Nasscom report on bridging the AI talent gap projects India's AI talent demand to exceed 1.25 million by 2027, against a base of roughly 600,000–650,000 in 2022, and found only about 16% of IT professionals were AI-skilled. When demand roughly doubles and the qualified pool stays thin, the clearing price moves faster than any annual cycle can track. So a band built from figures gathered 2 or 3 quarters ago is describing a market that has already moved on.

The second distortion is quieter than the lag and harder to spot, because it hides inside the word “average.” A benchmark is only as relevant as its sample, and most GCC surveys are dominated by whatever roles carry the highest volume, usually mainstream engineering. When you price a scarce, specialised seat against a basket built from abundant, generalist ones, the mix is working against you before you read a single number. The median you're quoted is the median of a population your candidate isn't in.

There's a self-selection effect underneath the mix, too. Surveys are populated by employers stable enough to participate, submitting roles that map cleanly to an established code. The mandate that's genuinely re-pricing a role, a new title, a hybrid skill set, a competitive counter-offer, is the one least likely to be sitting inside the sample. So the survey reports the market late and under-represents the very transactions setting the new price. The roles that would pull the average toward reality are the ones the sample excludes. That's why a benchmark can be entirely accurate about its basket and still mislead you about your seat: it answered a question about a different population and handed you the answer as if it were about yours.

None of this makes benchmarks useless. A good one bounds the conversation, keeps a committee honest, and gives you the most defensible starting point you have. The convergence between the major houses is itself a signal worth trusting: Aon projects India salary increases of 9.1% for 2026 and Mercer projects around 9%, two independent, gold-standard surveys landing in the same place. Treat that consensus as the base of the range. It tells you where the broad market sits; it doesn't tell you what the person in front of you will sign.

Correcting the band is a discipline. Start from the published midpoint, then age it forward by the drift you can actually observe: the offers your own recent mandates in that exact segment closed at, the counter-offers candidates are declining, and the leading indicators, live postings and movement in the senior salary band, that run ahead of any survey. Where the segment is re-pricing, weight it further; GCC-specific increments have run hotter than the national number, with Zinnov putting average GCC increments near 9.9%. And correct on the second axis too: adjust for mix as well as level, because a base-salary band tells you nothing about the variable and equity that close a senior candidate. The output is a working range for this role, this quarter, in this segment: the published band as a floor, corrected upward by the evidence you can observe.

For a compensation committee, all of this resolves into a short, repeatable routine. Before a senior band is signed off, ask five questions in order. When was this data collected, and how many quarters of lag am I importing? Which roles dominate the basket, and is my seat one of them or a scarce exception the median doesn't describe? Is this segment re-pricing, and if it's AI, data or a scarce IT skill, it almost certainly is? Is the number base-only or total-package, mix included? And finally: what did our own last comparable mandates actually close at, and how far above the published midpoint?

A committee that runs those five questions is reading the method before the number, and budgeting against what candidates are actually signing. A committee that reads the band as the answer is budgeting against a market that has already moved, on the roles it can least afford to under-price. Treating the benchmark as a verdict is the real mistake here. Read it as a summary of a known method, correct it for the two distortions that method encodes, and it becomes what it was always meant to be: a floor you build up from.