45 direct answers to the most-searched questions about applicant tracking systems, resume optimization, keyword strategy, and AI-written resumes in 2026. No marketing fluff.
Yes, but parsing accuracy varies. Workday, Greenhouse, and Lever can read most text-based PDFs cleanly. Image-rendered PDFs (scanned resumes, designer-exported PDFs with rasterised text) often fail to parse — the ATS sees blank pages. Test your PDF by copy-pasting from it; if the text comes out scrambled, your ATS will see scrambled text too. DOCX is the safer default in 2026.
Keyword stuffing means repeating JD keywords beyond what's natural — for example, putting 'Python developer Python engineer Python expert Python coder' in your skills section. Modern ATS in 2026 detect this via n-gram repetition checks and BERT-based semantic outlier detection. Stuffed resumes either get flagged for human review (bad if no recruiter checks) or auto-rejected outright.
Workday parses resumes into structured fields using a combination of rule-based extractors (regex for emails, phone numbers, dates) and a trained NER model for skills and job titles. The parsed structure is then matched against the JD's required-skills list with TF-IDF and embedding-similarity scores. Workday weights the SKILLS section higher than narrative paragraphs.
Google uses Avature for general roles and an internal system (sometimes called gHire) for engineering. Both are less keyword-aggressive than Workday — they rely more on human recruiter screening. But Google's volume is so high that a recruiter still spends 6-15 seconds per resume, so the same scannability principles apply.
Two-column resume layouts are a known parsing failure mode for several ATS (Taleo, older iCIMS configs). The parser reads left-to-right and merges the two columns into one stream, scrambling section order. Single-column layouts are the safe default. Tables for ranges (Jan 2020 — Mar 2022) usually parse fine; tables for whole sections do not.
One page if you have under 8 years of experience; two pages if you have more. Three+ pages for academic CVs and very senior roles only. ATS doesn't care about page count, but human recruiters skim — anything past page 2 has near-zero read probability for most roles.
Inter, Calibri, Arial, Helvetica, or Georgia. Avoid display fonts (Pacifico, Lobster, anything 'designer'), light weights, and embedded webfonts that aren't in the parser's font table. 10-11pt body, 14-18pt headings.
Not for US, UK, Canadian, or Australian roles — anti-bias regulations and recruiter convention reject photo'd resumes. Continental Europe (especially Germany) historically expected photos but is moving away. Asian markets remain mixed. When in doubt, skip the photo.
No. They were useful when resumes were career-shifts focused; today, the JD-targeted summary or 'Profile' paragraph replaces them. A four-line summary stating your level, two specialisations, and most-recent impact is what every modern template puts first.
In the US, a resume is 1-2 pages and tailored per role; a CV is the longer academic document with publications, talks, and full work history. In the UK, India, and most of Europe, 'CV' is used for what Americans call a resume — 1-2 pages, targeted. Match the regional convention of the company you're applying to.
Companies vary. FAANG and most large companies cross-check; discrepancies between LinkedIn and resume — different titles, different dates — get flagged. They don't need to be identical, but the rough shape (titles, employers, approximate dates) should align. Tailor the resume; keep LinkedIn as the canonical truth.
Three signals matter: (1) explicit timezone availability near the contact info, (2) async-collaboration evidence ('led 8-engineer distributed team across 4 timezones'), and (3) prior remote experience listed clearly per role ('Acme Corp, fully remote, 2020-2024'). Remote-job ATS queries filter for these literal phrases.
AI is great for *rewriting* sections you've already drafted; less good for generating from scratch. Recruiters in 2026 spot ChatGPT's signature phrases (utilised, leveraged, spearheaded, em-dashes) within seconds. Use AI as a sparring partner: write your bullet, ask it to sharpen, then rewrite in your own voice. Local LLMs (Llama-3.1) produce less homogenous output than ChatGPT.
Match the top 15-20 keywords from the specific JD; aim for 65-85% coverage. Beyond that you're keyword-stuffing. Track the rank order of keywords in the JD (first-mentioned = most important to the hiring manager) and mirror that ordering in your skills section.
Sparingly. One accent colour for headings or section dividers is fine. Background colours, coloured body text, and rainbow skill bars look amateur. Black-on-white text with one optional accent (navy, dark green, deep red) parses cleanly and prints well.
Reverse-chronological with: (1) 3-line profile summary, (2) experience with outcome-bullets (cut p99 from X to Y), (3) skills ordered by JD-relevance, not alphabetical, (4) education last for ICs with 5+ years experience. Skip 'objective', 'references available on request', and exhaustive certifications. DOCX format.
Mixed. High-volume engineering at FAANG: rarely. Senior IC and design roles: often. Mission-driven orgs (non-profit, edu, gov): consistently. Small companies (under 50 employees): always. When unsure, write one — the cost is 20 minutes; the upside on the 30% of orgs that read them is significant.
List them only when they're either (a) directly required by the JD ('AWS Solutions Architect required'), or (b) evidence of recent relevant learning (a 2025 cert is more useful than a 2018 one). Buried bottom of resume in a compact 'Certifications' line. Don't waste a full section on a single Coursera certificate.
Compress older roles. Years 0-3: title, company, dates only. Years 4-10: title, company, dates, 1-2 bullets. Most-recent 3 years: title, company, dates, 4-5 outcome bullets. Recruiters skim recent first; depth before breadth communicates seniority.
Most common causes (in order): (1) keyword match too low — use a free match-score tool to check, (2) PDF rendering issues — open it after upload to confirm parsing, (3) experience-level mismatch — applying to senior with junior framing, (4) location filter — many ATS auto-reject out-of-region candidates, (5) gap years not addressed explicitly.
Five concrete things: (1) text-extractable PDF or DOCX, (2) standard section headings ('Experience', not 'My Journey'), (3) single-column layout, (4) no embedded images for skills/headers, (5) skills explicitly listed in a dedicated 'Skills' or 'Technical Skills' section. Beyond those, 'ATS-friendly' is mostly marketing fluff.
Honestly and concisely. 'Career break, March 2023 – January 2024 — caregiver responsibilities, returned full-time.' Or 'Sabbatical, Jul 2022 – Feb 2023 — open-source contributions, see github.com/...'. Recruiters increasingly expect non-linear careers; what they reject is unexplained gaps.
Yes for the top 5-10 roles you actually want; no for the rest. Use a base template and rewrite the summary + skills section per JD; the experience bullets stay mostly stable. A tailored resume on average matches 40-50% better than a generic one — the question is whether the time investment fits your application volume.
Hard skills are verifiable technical competencies (Python, Kubernetes, financial modelling). Soft skills are behavioural traits (leadership, communication). ATS systems weight hard skills 5-10× higher because they match deterministically. List soft skills only when the JD explicitly demands them, and back each with a concrete example bullet, not as standalone words.
Some tools claim to (Originality.ai, GPTZero) but reliably detecting LLM-written prose remains an open problem. What ATS *can* detect is the homogeneity that emerges when everyone uses the same model: identical phrasing patterns across thousands of resumes look like a template. Variety in your phrasing matters more than 'beating' detection.
75-90% is the sweet spot. Below 60% suggests the role is a stretch or your resume needs tailoring. Above 90% looks suspiciously over-optimised — recruiters notice when every bullet contains JD vocab. Aim for the band that signals fit without signalling keyword-stuffing.
By default, yes — recordings, transcripts, and metadata are retained on Otter's AWS infrastructure. Their training opt-out exists but applies prospectively, not to data already collected. For sales calls, performance reviews, or any conversation with confidential business content, the default settings retain it indefinitely until you manually delete each recording.
Fireflies stores your meeting recordings on their cloud infrastructure and uses third-party sub-processors (listed in their DPA). 'Private' depends on your threat model — for external sharing, they're industry-standard; for competitive-intel concerns or regulated data, the answer is no. Self-hosted alternatives like locally-run Whisper process the audio without any cloud upload.
Yes. OpenAI Whisper, faster-whisper, and Whisper.cpp all run on Apple Silicon. On an M4 Pro, Whisper-base runs 5-10× faster than real-time on CPU, and Whisper-small (244M params) at 3-5×. No cloud round-trip, no API keys, audio never leaves the box. The catch: setup involves installing whisper-cli + ffmpeg + a model file.
On clean English audio, Whisper-large (local) matches or beats most cloud APIs (Deepgram, AssemblyAI). Where cloud wins: telephony audio, very heavy accents, real-time streaming with sub-300ms latency, and speaker diarization with PII redaction. For 80% of meeting/lecture/podcast use cases, local Whisper is competitive and free of recurring cost.
Both are local-runnable. Whisper-base (74M params) is faster (~10× real-time on M-series) with WER (word error rate) around 9-12% on clean audio. Whisper-small (244M params) is slower (~3-5×) but cleaner (~7-9% WER). For meeting notes where you'll re-read, base is fine; for podcast/lecture publishing, small is worth the extra time.
Local Whisper-base: about 6 min for 60 min of audio on an M-series Mac. Cloud APIs (Otter, Fireflies): typically real-time as the meeting happens, then 30s-2min for post-processing summary. Self-hosted on consumer hardware is slower than cloud but private. Cloud is faster but your audio touches their servers.
Transcript accuracy: 88-95% on clean audio (varies by speaker accent, jargon, audio quality). Summary accuracy: less measurable, but LLM-generated action items are *usually* on point if the meeting actually had clear action items. The failure mode is fabrication — LLMs occasionally invent commitments that nobody made. Always review the action-items list before sharing.
Speaker diarization (knowing which speaker said what) is a separate task from transcription. Cloud tools (Otter, Fireflies) ship with it built in. Local stacks need a diarization library (pyannote-audio is the open-source standard) layered on top of Whisper. Quality is decent for 2-4 speakers in clean audio; degrades fast with overlap or far-mic recordings.
Three rough options ranked by privacy: (1) self-hosted Whisper + LLM (audio never leaves your box); (2) on-device transcription apps that process locally; (3) cloud tools with explicit opt-out + DPA + business-tier encryption. Most enterprise concerns are addressed by option 3, but option 1 is the only one where the audio is cryptographically guaranteed not to leave your hardware.
A 60-minute meeting transcript is typically 8,000-12,000 words. Quality summaries land in 5-7 sentences for the executive summary, 4-8 action items, and 5-10 key points. Llama-3.1 8B locally produces this in about 30-60 seconds. Anything dramatically longer is the LLM padding — re-prompt for terseness.
Partly. AI tools (Tome, Beautiful.ai, Gamma, local-LLM alternatives) draft an outline + design well in 2026. What they CAN'T do: capture your founder voice, your specific competitive advantage, or your defensibility thesis. Use AI for the structural skeleton (10-12 slides in standard order) + your own writing for the 2-3 slides that actually decide the meeting (ask, traction, why-now).
Tome is a HTML-first canvas — slides live in their viewer, PDF export is paywalled at $20/mo. Beautiful.ai is more design-system-driven (auto-layout). Both retain your topic + outline on their cloud. Local-LLM alternatives like ANANTA Slides skip the cloud upload entirely and export PDF directly.
10-12 for a seed round (problem, solution, market, traction, business model, competition, team, ask). 15-20 for Series A. Anything beyond 20 means you haven't decided what matters. Indian/US VCs both spend 3-4 minutes on a deck on first pass — design for skim-readability.
Yes — Tome, Beautiful.ai, Gamma all have free tiers (usually 3-5 decks/month with their watermark). ANANTA Slides offers 3 free decks/month with PDF export and no watermark, since it runs on a Mac Mini that's already amortised. The trade-off is design flexibility — local-LLM tools have fewer template options than the cloud incumbents.
For seed-round decks, internal status updates, and student presentations: yes, with edits. For Series A/B and external fundraising: AI gets you to a 60-70% draft in 20 seconds; human polish for the 30-40% that matters most. The anti-pattern is shipping AI output verbatim — investors spot the patterns within seconds.
Few tools export to .pptx natively (Beautiful.ai is one). Most export to PDF or PNG. The .pptx ecosystem is brittle (font substitution, shape rendering varies between Mac and Windows). PDF is the safer default for sharing — fonts embedded, layout fixed, no edit-permissions confusion.
Outline-only tools (just text): under 5 seconds with a small LLM. Design-rendered decks: 15-60 seconds depending on slide count and renderer. ANANTA Slides produces an 8-slide PDF in ~15 seconds; a 20-slide deck in ~25 seconds (Llama-3.1 8B + Playwright Chromium on M4 Pro).
Depends on the tool's data policy. Most cloud AI deck tools retain your input + drafts on their infrastructure for model training (opt-out exists, defaults are on). For pre-launch / stealth ideas, this is a real risk. Self-hosted / local-LLM tools (ANANTA Slides) don't send your topic anywhere; outline + render happen on a single Mac Mini that nobody else has access to.
Some tools support 'document → deck' (Tome, Gamma). The quality depends on the document — well-structured docs (headers, sections) produce better decks than long prose. For meeting notes → deck, the workflow is: transcript → structured action items → slide outline. ANANTA Notes + ANANTA Slides chain together for this; alternatively just feed the meeting summary into Slides directly.
Stop guessing about ATS scoring — measure your resume against any JD:
ATS keyword extractor → Resume vs JD match score →